1 /* 2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/subnode.hpp" 43 #include "opto/valuetypenode.hpp" 44 #include "utilities/macros.hpp" 45 #if INCLUDE_ZGC 46 #include "gc/z/c2/zBarrierSetC2.hpp" 47 #endif 48 49 //============================================================================= 50 //------------------------------split_thru_phi--------------------------------- 51 // Split Node 'n' through merge point if there is enough win. 52 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 53 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 54 // ConvI2L may have type information on it which is unsafe to push up 55 // so disable this for now 56 return NULL; 57 } 58 59 // Splitting range check CastIIs through a loop induction Phi can 60 // cause new Phis to be created that are left unrelated to the loop 61 // induction Phi and prevent optimizations (vectorization) 62 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 63 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 64 return NULL; 65 } 66 67 // Value types should not be split through Phis because they cannot be merged 68 // through Phi nodes but each value input needs to be merged individually. 69 if (n->is_ValueType()) { 70 return NULL; 71 } 72 73 int wins = 0; 74 assert(!n->is_CFG(), ""); 75 assert(region->is_Region(), ""); 76 77 const Type* type = n->bottom_type(); 78 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 79 Node *phi; 80 if (t_oop != NULL && t_oop->is_known_instance_field()) { 81 int iid = t_oop->instance_id(); 82 int index = C->get_alias_index(t_oop); 83 int offset = t_oop->offset(); 84 phi = new PhiNode(region, type, NULL, iid, index, offset); 85 } else { 86 phi = PhiNode::make_blank(region, n); 87 } 88 uint old_unique = C->unique(); 89 for (uint i = 1; i < region->req(); i++) { 90 Node *x; 91 Node* the_clone = NULL; 92 if (region->in(i) == C->top()) { 93 x = C->top(); // Dead path? Use a dead data op 94 } else { 95 x = n->clone(); // Else clone up the data op 96 the_clone = x; // Remember for possible deletion. 97 // Alter data node to use pre-phi inputs 98 if (n->in(0) == region) 99 x->set_req( 0, region->in(i) ); 100 for (uint j = 1; j < n->req(); j++) { 101 Node *in = n->in(j); 102 if (in->is_Phi() && in->in(0) == region) 103 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 104 } 105 } 106 // Check for a 'win' on some paths 107 const Type *t = x->Value(&_igvn); 108 109 bool singleton = t->singleton(); 110 111 // A TOP singleton indicates that there are no possible values incoming 112 // along a particular edge. In most cases, this is OK, and the Phi will 113 // be eliminated later in an Ideal call. However, we can't allow this to 114 // happen if the singleton occurs on loop entry, as the elimination of 115 // the PhiNode may cause the resulting node to migrate back to a previous 116 // loop iteration. 117 if (singleton && t == Type::TOP) { 118 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 119 // irreducible loop may not be indicated by an affirmative is_Loop()); 120 // therefore, the only top we can split thru a phi is on a backedge of 121 // a loop. 122 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 123 } 124 125 if (singleton) { 126 wins++; 127 x = ((PhaseGVN&)_igvn).makecon(t); 128 } else { 129 // We now call Identity to try to simplify the cloned node. 130 // Note that some Identity methods call phase->type(this). 131 // Make sure that the type array is big enough for 132 // our new node, even though we may throw the node away. 133 // (Note: This tweaking with igvn only works because x is a new node.) 134 _igvn.set_type(x, t); 135 // If x is a TypeNode, capture any more-precise type permanently into Node 136 // otherwise it will be not updated during igvn->transform since 137 // igvn->type(x) is set to x->Value() already. 138 x->raise_bottom_type(t); 139 Node *y = _igvn.apply_identity(x); 140 if (y != x) { 141 wins++; 142 x = y; 143 } else { 144 y = _igvn.hash_find(x); 145 if (y) { 146 wins++; 147 x = y; 148 } else { 149 // Else x is a new node we are keeping 150 // We do not need register_new_node_with_optimizer 151 // because set_type has already been called. 152 _igvn._worklist.push(x); 153 } 154 } 155 } 156 if (x != the_clone && the_clone != NULL) 157 _igvn.remove_dead_node(the_clone); 158 phi->set_req( i, x ); 159 } 160 // Too few wins? 161 if (wins <= policy) { 162 _igvn.remove_dead_node(phi); 163 return NULL; 164 } 165 166 // Record Phi 167 register_new_node( phi, region ); 168 169 for (uint i2 = 1; i2 < phi->req(); i2++) { 170 Node *x = phi->in(i2); 171 // If we commoned up the cloned 'x' with another existing Node, 172 // the existing Node picks up a new use. We need to make the 173 // existing Node occur higher up so it dominates its uses. 174 Node *old_ctrl; 175 IdealLoopTree *old_loop; 176 177 if (x->is_Con()) { 178 // Constant's control is always root. 179 set_ctrl(x, C->root()); 180 continue; 181 } 182 // The occasional new node 183 if (x->_idx >= old_unique) { // Found a new, unplaced node? 184 old_ctrl = NULL; 185 old_loop = NULL; // Not in any prior loop 186 } else { 187 old_ctrl = get_ctrl(x); 188 old_loop = get_loop(old_ctrl); // Get prior loop 189 } 190 // New late point must dominate new use 191 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 192 if (new_ctrl == old_ctrl) // Nothing is changed 193 continue; 194 195 IdealLoopTree *new_loop = get_loop(new_ctrl); 196 197 // Don't move x into a loop if its uses are 198 // outside of loop. Otherwise x will be cloned 199 // for each use outside of this loop. 200 IdealLoopTree *use_loop = get_loop(region); 201 if (!new_loop->is_member(use_loop) && 202 (old_loop == NULL || !new_loop->is_member(old_loop))) { 203 // Take early control, later control will be recalculated 204 // during next iteration of loop optimizations. 205 new_ctrl = get_early_ctrl(x); 206 new_loop = get_loop(new_ctrl); 207 } 208 // Set new location 209 set_ctrl(x, new_ctrl); 210 // If changing loop bodies, see if we need to collect into new body 211 if (old_loop != new_loop) { 212 if (old_loop && !old_loop->_child) 213 old_loop->_body.yank(x); 214 if (!new_loop->_child) 215 new_loop->_body.push(x); // Collect body info 216 } 217 } 218 219 return phi; 220 } 221 222 //------------------------------dominated_by------------------------------------ 223 // Replace the dominated test with an obvious true or false. Place it on the 224 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 225 // live path up to the dominating control. 226 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 227 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 228 229 // prevdom is the dominating projection of the dominating test. 230 assert( iff->is_If(), "" ); 231 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 232 int pop = prevdom->Opcode(); 233 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 234 if (flip) { 235 if (pop == Op_IfTrue) 236 pop = Op_IfFalse; 237 else 238 pop = Op_IfTrue; 239 } 240 // 'con' is set to true or false to kill the dominated test. 241 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 242 set_ctrl(con, C->root()); // Constant gets a new use 243 // Hack the dominated test 244 _igvn.replace_input_of(iff, 1, con); 245 246 // If I dont have a reachable TRUE and FALSE path following the IfNode then 247 // I can assume this path reaches an infinite loop. In this case it's not 248 // important to optimize the data Nodes - either the whole compilation will 249 // be tossed or this path (and all data Nodes) will go dead. 250 if (iff->outcnt() != 2) return; 251 252 // Make control-dependent data Nodes on the live path (path that will remain 253 // once the dominated IF is removed) become control-dependent on the 254 // dominating projection. 255 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 256 257 // Loop predicates may have depending checks which should not 258 // be skipped. For example, range check predicate has two checks 259 // for lower and upper bounds. 260 if (dp == NULL) 261 return; 262 263 ProjNode* dp_proj = dp->as_Proj(); 264 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 265 if (exclude_loop_predicate && 266 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 267 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 268 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 269 // If this is a range check (IfNode::is_range_check), do not 270 // reorder because Compile::allow_range_check_smearing might have 271 // changed the check. 272 return; // Let IGVN transformation change control dependence. 273 } 274 275 IdealLoopTree *old_loop = get_loop(dp); 276 277 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 278 Node* cd = dp->fast_out(i); // Control-dependent node 279 if (cd->depends_only_on_test()) { 280 assert(cd->in(0) == dp, ""); 281 _igvn.replace_input_of(cd, 0, prevdom); 282 set_early_ctrl(cd); 283 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 284 if (old_loop != new_loop) { 285 if (!old_loop->_child) old_loop->_body.yank(cd); 286 if (!new_loop->_child) new_loop->_body.push(cd); 287 } 288 --i; 289 --imax; 290 } 291 } 292 } 293 294 //------------------------------has_local_phi_input---------------------------- 295 // Return TRUE if 'n' has Phi inputs from its local block and no other 296 // block-local inputs (all non-local-phi inputs come from earlier blocks) 297 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 298 Node *n_ctrl = get_ctrl(n); 299 // See if some inputs come from a Phi in this block, or from before 300 // this block. 301 uint i; 302 for( i = 1; i < n->req(); i++ ) { 303 Node *phi = n->in(i); 304 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 305 break; 306 } 307 if( i >= n->req() ) 308 return NULL; // No Phi inputs; nowhere to clone thru 309 310 // Check for inputs created between 'n' and the Phi input. These 311 // must split as well; they have already been given the chance 312 // (courtesy of a post-order visit) and since they did not we must 313 // recover the 'cost' of splitting them by being very profitable 314 // when splitting 'n'. Since this is unlikely we simply give up. 315 for( i = 1; i < n->req(); i++ ) { 316 Node *m = n->in(i); 317 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 318 // We allow the special case of AddP's with no local inputs. 319 // This allows us to split-up address expressions. 320 if (m->is_AddP() && 321 get_ctrl(m->in(2)) != n_ctrl && 322 get_ctrl(m->in(3)) != n_ctrl) { 323 // Move the AddP up to dominating point 324 Node* c = find_non_split_ctrl(idom(n_ctrl)); 325 if (c->is_OuterStripMinedLoop()) { 326 c->as_Loop()->verify_strip_mined(1); 327 c = c->in(LoopNode::EntryControl); 328 } 329 set_ctrl_and_loop(m, c); 330 continue; 331 } 332 return NULL; 333 } 334 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 335 } 336 337 return n_ctrl; 338 } 339 340 //------------------------------remix_address_expressions---------------------- 341 // Rework addressing expressions to get the most loop-invariant stuff 342 // moved out. We'd like to do all associative operators, but it's especially 343 // important (common) to do address expressions. 344 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 345 if (!has_ctrl(n)) return NULL; 346 Node *n_ctrl = get_ctrl(n); 347 IdealLoopTree *n_loop = get_loop(n_ctrl); 348 349 // See if 'n' mixes loop-varying and loop-invariant inputs and 350 // itself is loop-varying. 351 352 // Only interested in binary ops (and AddP) 353 if( n->req() < 3 || n->req() > 4 ) return NULL; 354 355 Node *n1_ctrl = get_ctrl(n->in( 1)); 356 Node *n2_ctrl = get_ctrl(n->in( 2)); 357 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 358 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 359 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 360 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 361 362 // Does one of my inputs spin in a tighter loop than self? 363 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 364 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 365 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 366 return NULL; // Leave well enough alone 367 368 // Is at least one of my inputs loop-invariant? 369 if( n1_loop == n_loop && 370 n2_loop == n_loop && 371 n3_loop == n_loop ) 372 return NULL; // No loop-invariant inputs 373 374 375 int n_op = n->Opcode(); 376 377 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 378 if( n_op == Op_LShiftI ) { 379 // Scale is loop invariant 380 Node *scale = n->in(2); 381 Node *scale_ctrl = get_ctrl(scale); 382 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 383 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 384 return NULL; 385 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 386 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 387 return NULL; // Dont bother with byte/short masking 388 // Add must vary with loop (else shift would be loop-invariant) 389 Node *add = n->in(1); 390 Node *add_ctrl = get_ctrl(add); 391 IdealLoopTree *add_loop = get_loop(add_ctrl); 392 //assert( n_loop == add_loop, "" ); 393 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 394 395 // Convert I-V into I+ (0-V); same for V-I 396 if( add->Opcode() == Op_SubI && 397 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 398 Node *zero = _igvn.intcon(0); 399 set_ctrl(zero, C->root()); 400 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 401 register_new_node( neg, get_ctrl(add->in(2) ) ); 402 add = new AddINode( add->in(1), neg ); 403 register_new_node( add, add_ctrl ); 404 } 405 if( add->Opcode() != Op_AddI ) return NULL; 406 // See if one add input is loop invariant 407 Node *add_var = add->in(1); 408 Node *add_var_ctrl = get_ctrl(add_var); 409 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 410 Node *add_invar = add->in(2); 411 Node *add_invar_ctrl = get_ctrl(add_invar); 412 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 413 if( add_var_loop == n_loop ) { 414 } else if( add_invar_loop == n_loop ) { 415 // Swap to find the invariant part 416 add_invar = add_var; 417 add_invar_ctrl = add_var_ctrl; 418 add_invar_loop = add_var_loop; 419 add_var = add->in(2); 420 Node *add_var_ctrl = get_ctrl(add_var); 421 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 422 } else // Else neither input is loop invariant 423 return NULL; 424 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 425 return NULL; // No invariant part of the add? 426 427 // Yes! Reshape address expression! 428 Node *inv_scale = new LShiftINode( add_invar, scale ); 429 Node *inv_scale_ctrl = 430 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 431 add_invar_ctrl : scale_ctrl; 432 register_new_node( inv_scale, inv_scale_ctrl ); 433 Node *var_scale = new LShiftINode( add_var, scale ); 434 register_new_node( var_scale, n_ctrl ); 435 Node *var_add = new AddINode( var_scale, inv_scale ); 436 register_new_node( var_add, n_ctrl ); 437 _igvn.replace_node( n, var_add ); 438 return var_add; 439 } 440 441 // Replace (I+V) with (V+I) 442 if( n_op == Op_AddI || 443 n_op == Op_AddL || 444 n_op == Op_AddF || 445 n_op == Op_AddD || 446 n_op == Op_MulI || 447 n_op == Op_MulL || 448 n_op == Op_MulF || 449 n_op == Op_MulD ) { 450 if( n2_loop == n_loop ) { 451 assert( n1_loop != n_loop, "" ); 452 n->swap_edges(1, 2); 453 } 454 } 455 456 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 457 // but not if I2 is a constant. 458 if( n_op == Op_AddP ) { 459 if( n2_loop == n_loop && n3_loop != n_loop ) { 460 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 461 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 462 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 463 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 464 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 465 if( n22loop != n_loop && n22loop->is_member(n_loop) && 466 n23_loop == n_loop ) { 467 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 468 // Stuff new AddP in the loop preheader 469 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 470 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 471 register_new_node( add2, n_ctrl ); 472 _igvn.replace_node( n, add2 ); 473 return add2; 474 } 475 } 476 } 477 478 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 479 if (n2_loop != n_loop && n3_loop == n_loop) { 480 if (n->in(3)->Opcode() == Op_AddX) { 481 Node *V = n->in(3)->in(1); 482 Node *I = n->in(3)->in(2); 483 if (is_member(n_loop,get_ctrl(V))) { 484 } else { 485 Node *tmp = V; V = I; I = tmp; 486 } 487 if (!is_member(n_loop,get_ctrl(I))) { 488 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 489 // Stuff new AddP in the loop preheader 490 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 491 Node *add2 = new AddPNode(n->in(1), add1, V); 492 register_new_node(add2, n_ctrl); 493 _igvn.replace_node(n, add2); 494 return add2; 495 } 496 } 497 } 498 } 499 500 return NULL; 501 } 502 503 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 504 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 505 assert(n->Opcode() == Op_AddI, "sanity"); 506 Node * nn = NULL; 507 Node * in1 = n->in(1); 508 Node * in2 = n->in(2); 509 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 510 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 511 if (loop_n->_head->as_Loop()->is_valid_counted_loop() && 512 Matcher::match_rule_supported(Op_MulAddS2I) && 513 Matcher::match_rule_supported(Op_MulAddVS2VI)) { 514 Node* mul_in1 = in1->in(1); 515 Node* mul_in2 = in1->in(2); 516 Node* mul_in3 = in2->in(1); 517 Node* mul_in4 = in2->in(2); 518 if (mul_in1->Opcode() == Op_LoadS && 519 mul_in2->Opcode() == Op_LoadS && 520 mul_in3->Opcode() == Op_LoadS && 521 mul_in4->Opcode() == Op_LoadS) { 522 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 523 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 524 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 525 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 526 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 527 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 528 // All nodes should be in the same counted loop. 529 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 530 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 531 Node* adr1 = mul_in1->in(MemNode::Address); 532 Node* adr2 = mul_in2->in(MemNode::Address); 533 Node* adr3 = mul_in3->in(MemNode::Address); 534 Node* adr4 = mul_in4->in(MemNode::Address); 535 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 536 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 537 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 538 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 539 register_new_node(nn, get_ctrl(n)); 540 _igvn.replace_node(n, nn); 541 return nn; 542 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 543 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 544 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 545 register_new_node(nn, get_ctrl(n)); 546 _igvn.replace_node(n, nn); 547 return nn; 548 } 549 } 550 } 551 } 552 } 553 } 554 return nn; 555 } 556 557 //------------------------------conditional_move------------------------------- 558 // Attempt to replace a Phi with a conditional move. We have some pretty 559 // strict profitability requirements. All Phis at the merge point must 560 // be converted, so we can remove the control flow. We need to limit the 561 // number of c-moves to a small handful. All code that was in the side-arms 562 // of the CFG diamond is now speculatively executed. This code has to be 563 // "cheap enough". We are pretty much limited to CFG diamonds that merge 564 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 565 Node *PhaseIdealLoop::conditional_move( Node *region ) { 566 567 assert(region->is_Region(), "sanity check"); 568 if (region->req() != 3) return NULL; 569 570 // Check for CFG diamond 571 Node *lp = region->in(1); 572 Node *rp = region->in(2); 573 if (!lp || !rp) return NULL; 574 Node *lp_c = lp->in(0); 575 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 576 IfNode *iff = lp_c->as_If(); 577 578 // Check for ops pinned in an arm of the diamond. 579 // Can't remove the control flow in this case 580 if (lp->outcnt() > 1) return NULL; 581 if (rp->outcnt() > 1) return NULL; 582 583 IdealLoopTree* r_loop = get_loop(region); 584 assert(r_loop == get_loop(iff), "sanity"); 585 // Always convert to CMOVE if all results are used only outside this loop. 586 bool used_inside_loop = (r_loop == _ltree_root); 587 588 // Check profitability 589 int cost = 0; 590 int phis = 0; 591 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 592 Node *out = region->fast_out(i); 593 if (!out->is_Phi()) continue; // Ignore other control edges, etc 594 phis++; 595 PhiNode* phi = out->as_Phi(); 596 BasicType bt = phi->type()->basic_type(); 597 switch (bt) { 598 case T_DOUBLE: 599 case T_FLOAT: 600 if (C->use_cmove()) { 601 continue; //TODO: maybe we want to add some cost 602 } 603 cost += Matcher::float_cmove_cost(); // Could be very expensive 604 break; 605 case T_LONG: { 606 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 607 } 608 case T_INT: // These all CMOV fine 609 case T_ADDRESS: { // (RawPtr) 610 cost++; 611 break; 612 } 613 case T_NARROWOOP: // Fall through 614 case T_OBJECT: { // Base oops are OK, but not derived oops 615 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 616 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 617 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 618 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 619 // have a Phi for the base here that we convert to a CMOVE all is well 620 // and good. But if the base is dead, we'll not make a CMOVE. Later 621 // the allocator will have to produce a base by creating a CMOVE of the 622 // relevant bases. This puts the allocator in the business of 623 // manufacturing expensive instructions, generally a bad plan. 624 // Just Say No to Conditionally-Moved Derived Pointers. 625 if (tp && tp->offset() != 0) 626 return NULL; 627 cost++; 628 break; 629 } 630 default: 631 return NULL; // In particular, can't do memory or I/O 632 } 633 // Add in cost any speculative ops 634 for (uint j = 1; j < region->req(); j++) { 635 Node *proj = region->in(j); 636 Node *inp = phi->in(j); 637 if (get_ctrl(inp) == proj) { // Found local op 638 cost++; 639 // Check for a chain of dependent ops; these will all become 640 // speculative in a CMOV. 641 for (uint k = 1; k < inp->req(); k++) 642 if (get_ctrl(inp->in(k)) == proj) 643 cost += ConditionalMoveLimit; // Too much speculative goo 644 } 645 } 646 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 647 // This will likely Split-If, a higher-payoff operation. 648 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 649 Node* use = phi->fast_out(k); 650 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 651 cost += ConditionalMoveLimit; 652 // Is there a use inside the loop? 653 // Note: check only basic types since CMoveP is pinned. 654 if (!used_inside_loop && is_java_primitive(bt)) { 655 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 656 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 657 used_inside_loop = true; 658 } 659 } 660 } 661 }//for 662 Node* bol = iff->in(1); 663 assert(bol->Opcode() == Op_Bool, ""); 664 int cmp_op = bol->in(1)->Opcode(); 665 // It is expensive to generate flags from a float compare. 666 // Avoid duplicated float compare. 667 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 668 669 float infrequent_prob = PROB_UNLIKELY_MAG(3); 670 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 671 if (used_inside_loop) { 672 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 673 674 // BlockLayoutByFrequency optimization moves infrequent branch 675 // from hot path. No point in CMOV'ing in such case (110 is used 676 // instead of 100 to take into account not exactness of float value). 677 if (BlockLayoutByFrequency) { 678 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 679 } 680 } 681 // Check for highly predictable branch. No point in CMOV'ing if 682 // we are going to predict accurately all the time. 683 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 684 //keep going 685 } else if (iff->_prob < infrequent_prob || 686 iff->_prob > (1.0f - infrequent_prob)) 687 return NULL; 688 689 // -------------- 690 // Now replace all Phis with CMOV's 691 Node *cmov_ctrl = iff->in(0); 692 uint flip = (lp->Opcode() == Op_IfTrue); 693 Node_List wq; 694 while (1) { 695 PhiNode* phi = NULL; 696 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 697 Node *out = region->fast_out(i); 698 if (out->is_Phi()) { 699 phi = out->as_Phi(); 700 break; 701 } 702 } 703 if (phi == NULL) break; 704 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 705 // Move speculative ops 706 wq.push(phi); 707 while (wq.size() > 0) { 708 Node *n = wq.pop(); 709 for (uint j = 1; j < n->req(); j++) { 710 Node* m = n->in(j); 711 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 712 #ifndef PRODUCT 713 if (PrintOpto && VerifyLoopOptimizations) { 714 tty->print(" speculate: "); 715 m->dump(); 716 } 717 #endif 718 set_ctrl(m, cmov_ctrl); 719 wq.push(m); 720 } 721 } 722 } 723 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 724 register_new_node( cmov, cmov_ctrl ); 725 _igvn.replace_node( phi, cmov ); 726 #ifndef PRODUCT 727 if (TraceLoopOpts) { 728 tty->print("CMOV "); 729 r_loop->dump_head(); 730 if (Verbose) { 731 bol->in(1)->dump(1); 732 cmov->dump(1); 733 } 734 } 735 if (VerifyLoopOptimizations) verify(); 736 #endif 737 } 738 739 // The useless CFG diamond will fold up later; see the optimization in 740 // RegionNode::Ideal. 741 _igvn._worklist.push(region); 742 743 return iff->in(1); 744 } 745 746 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 747 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 748 Node* u = m->fast_out(i); 749 if (u->is_CFG()) { 750 if (u->Opcode() == Op_NeverBranch) { 751 u = ((NeverBranchNode*)u)->proj_out(0); 752 enqueue_cfg_uses(u, wq); 753 } else { 754 wq.push(u); 755 } 756 } 757 } 758 } 759 760 // Try moving a store out of a loop, right before the loop 761 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 762 // Store has to be first in the loop body 763 IdealLoopTree *n_loop = get_loop(n_ctrl); 764 if (n->is_Store() && n_loop != _ltree_root && 765 n_loop->is_loop() && n_loop->_head->is_Loop() && 766 n->in(0) != NULL) { 767 Node* address = n->in(MemNode::Address); 768 Node* value = n->in(MemNode::ValueIn); 769 Node* mem = n->in(MemNode::Memory); 770 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 771 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 772 773 // - address and value must be loop invariant 774 // - memory must be a memory Phi for the loop 775 // - Store must be the only store on this memory slice in the 776 // loop: if there's another store following this one then value 777 // written at iteration i by the second store could be overwritten 778 // at iteration i+n by the first store: it's not safe to move the 779 // first store out of the loop 780 // - nothing must observe the memory Phi: it guarantees no read 781 // before the store, we are also guaranteed the store post 782 // dominates the loop head (ignoring a possible early 783 // exit). Otherwise there would be extra Phi involved between the 784 // loop's Phi and the store. 785 // - there must be no early exit from the loop before the Store 786 // (such an exit most of the time would be an extra use of the 787 // memory Phi but sometimes is a bottom memory Phi that takes the 788 // store as input). 789 790 if (!n_loop->is_member(address_loop) && 791 !n_loop->is_member(value_loop) && 792 mem->is_Phi() && mem->in(0) == n_loop->_head && 793 mem->outcnt() == 1 && 794 mem->in(LoopNode::LoopBackControl) == n) { 795 796 assert(n_loop->_tail != NULL, "need a tail"); 797 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 798 799 // Verify that there's no early exit of the loop before the store. 800 bool ctrl_ok = false; 801 { 802 // Follow control from loop head until n, we exit the loop or 803 // we reach the tail 804 ResourceMark rm; 805 Unique_Node_List wq; 806 wq.push(n_loop->_head); 807 808 for (uint next = 0; next < wq.size(); ++next) { 809 Node *m = wq.at(next); 810 if (m == n->in(0)) { 811 ctrl_ok = true; 812 continue; 813 } 814 assert(!has_ctrl(m), "should be CFG"); 815 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 816 ctrl_ok = false; 817 break; 818 } 819 enqueue_cfg_uses(m, wq); 820 if (wq.size() > 10) { 821 ctrl_ok = false; 822 break; 823 } 824 } 825 } 826 if (ctrl_ok) { 827 // move the Store 828 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 829 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 830 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 831 // Disconnect the phi now. An empty phi can confuse other 832 // optimizations in this pass of loop opts. 833 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 834 n_loop->_body.yank(mem); 835 836 set_ctrl_and_loop(n, n->in(0)); 837 838 return n; 839 } 840 } 841 } 842 return NULL; 843 } 844 845 // Try moving a store out of a loop, right after the loop 846 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 847 if (n->is_Store() && n->in(0) != NULL) { 848 Node *n_ctrl = get_ctrl(n); 849 IdealLoopTree *n_loop = get_loop(n_ctrl); 850 // Store must be in a loop 851 if (n_loop != _ltree_root && !n_loop->_irreducible) { 852 Node* address = n->in(MemNode::Address); 853 Node* value = n->in(MemNode::ValueIn); 854 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 855 // address must be loop invariant 856 if (!n_loop->is_member(address_loop)) { 857 // Store must be last on this memory slice in the loop and 858 // nothing in the loop must observe it 859 Node* phi = NULL; 860 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 861 Node* u = n->fast_out(i); 862 if (has_ctrl(u)) { // control use? 863 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 864 if (!n_loop->is_member(u_loop)) { 865 continue; 866 } 867 if (u->is_Phi() && u->in(0) == n_loop->_head) { 868 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 869 // multiple phis on the same slice are possible 870 if (phi != NULL) { 871 return; 872 } 873 phi = u; 874 continue; 875 } 876 } 877 return; 878 } 879 if (phi != NULL) { 880 // Nothing in the loop before the store (next iteration) 881 // must observe the stored value 882 bool mem_ok = true; 883 { 884 ResourceMark rm; 885 Unique_Node_List wq; 886 wq.push(phi); 887 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 888 Node *m = wq.at(next); 889 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 890 Node* u = m->fast_out(i); 891 if (u->is_Store() || u->is_Phi()) { 892 if (u != n) { 893 wq.push(u); 894 mem_ok = (wq.size() <= 10); 895 } 896 } else { 897 mem_ok = false; 898 break; 899 } 900 } 901 } 902 } 903 if (mem_ok) { 904 // Move the store out of the loop if the LCA of all 905 // users (except for the phi) is outside the loop. 906 Node* hook = new Node(1); 907 _igvn.rehash_node_delayed(phi); 908 int count = phi->replace_edge(n, hook); 909 assert(count > 0, "inconsistent phi"); 910 911 // Compute latest point this store can go 912 Node* lca = get_late_ctrl(n, get_ctrl(n)); 913 if (n_loop->is_member(get_loop(lca))) { 914 // LCA is in the loop - bail out 915 _igvn.replace_node(hook, n); 916 return; 917 } 918 #ifdef ASSERT 919 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 920 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 921 n_loop->_head->as_Loop()->verify_strip_mined(1); 922 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 923 IdealLoopTree* outer_loop = get_loop(outer); 924 assert(n_loop->_parent == outer_loop, "broken loop tree"); 925 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 926 } 927 #endif 928 929 // Move store out of the loop 930 _igvn.replace_node(hook, n->in(MemNode::Memory)); 931 _igvn.replace_input_of(n, 0, lca); 932 set_ctrl_and_loop(n, lca); 933 934 // Disconnect the phi now. An empty phi can confuse other 935 // optimizations in this pass of loop opts.. 936 if (phi->in(LoopNode::LoopBackControl) == phi) { 937 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 938 n_loop->_body.yank(phi); 939 } 940 } 941 } 942 } 943 } 944 } 945 } 946 947 //------------------------------split_if_with_blocks_pre----------------------- 948 // Do the real work in a non-recursive function. Data nodes want to be 949 // cloned in the pre-order so they can feed each other nicely. 950 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 952 Node* bs_res = bs->split_if_pre(this, n); 953 if (bs_res != NULL) { 954 return bs_res; 955 } 956 // Cloning these guys is unlikely to win 957 int n_op = n->Opcode(); 958 if( n_op == Op_MergeMem ) return n; 959 if( n->is_Proj() ) return n; 960 // Do not clone-up CmpFXXX variations, as these are always 961 // followed by a CmpI 962 if( n->is_Cmp() ) return n; 963 // Attempt to use a conditional move instead of a phi/branch 964 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 965 Node *cmov = conditional_move( n ); 966 if( cmov ) return cmov; 967 } 968 if( n->is_CFG() || n->is_LoadStore() ) 969 return n; 970 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 971 n_op == Op_Opaque2 ) { 972 if( !C->major_progress() ) // If chance of no more loop opts... 973 _igvn._worklist.push(n); // maybe we'll remove them 974 return n; 975 } 976 977 if( n->is_Con() ) return n; // No cloning for Con nodes 978 979 Node *n_ctrl = get_ctrl(n); 980 if( !n_ctrl ) return n; // Dead node 981 982 Node* res = try_move_store_before_loop(n, n_ctrl); 983 if (res != NULL) { 984 return n; 985 } 986 987 // Attempt to remix address expressions for loop invariants 988 Node *m = remix_address_expressions( n ); 989 if( m ) return m; 990 991 if (n_op == Op_AddI) { 992 Node *nn = convert_add_to_muladd( n ); 993 if ( nn ) return nn; 994 } 995 996 if (n->is_ConstraintCast()) { 997 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 998 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 999 // Node control inputs don't necessarily agree with loop control info (due to 1000 // transformations happened in between), thus additional dominance check is needed 1001 // to keep loop info valid. 1002 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1003 _igvn.replace_node(n, dom_cast); 1004 return dom_cast; 1005 } 1006 } 1007 1008 // Determine if the Node has inputs from some local Phi. 1009 // Returns the block to clone thru. 1010 Node *n_blk = has_local_phi_input( n ); 1011 if( !n_blk ) return n; 1012 1013 // Do not clone the trip counter through on a CountedLoop 1014 // (messes up the canonical shape). 1015 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1016 1017 // Check for having no control input; not pinned. Allow 1018 // dominating control. 1019 if (n->in(0)) { 1020 Node *dom = idom(n_blk); 1021 if (dom_lca(n->in(0), dom) != n->in(0)) { 1022 return n; 1023 } 1024 } 1025 // Policy: when is it profitable. You must get more wins than 1026 // policy before it is considered profitable. Policy is usually 0, 1027 // so 1 win is considered profitable. Big merges will require big 1028 // cloning, so get a larger policy. 1029 int policy = n_blk->req() >> 2; 1030 1031 // If the loop is a candidate for range check elimination, 1032 // delay splitting through it's phi until a later loop optimization 1033 if (n_blk->is_CountedLoop()) { 1034 IdealLoopTree *lp = get_loop(n_blk); 1035 if (lp && lp->_rce_candidate) { 1036 return n; 1037 } 1038 } 1039 1040 if (must_throttle_split_if()) return n; 1041 1042 // Split 'n' through the merge point if it is profitable 1043 Node *phi = split_thru_phi( n, n_blk, policy ); 1044 if (!phi) return n; 1045 1046 // Found a Phi to split thru! 1047 // Replace 'n' with the new phi 1048 _igvn.replace_node( n, phi ); 1049 // Moved a load around the loop, 'en-registering' something. 1050 if (n_blk->is_Loop() && n->is_Load() && 1051 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1052 C->set_major_progress(); 1053 1054 return phi; 1055 } 1056 1057 static bool merge_point_too_heavy(Compile* C, Node* region) { 1058 // Bail out if the region and its phis have too many users. 1059 int weight = 0; 1060 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1061 weight += region->fast_out(i)->outcnt(); 1062 } 1063 int nodes_left = C->max_node_limit() - C->live_nodes(); 1064 if (weight * 8 > nodes_left) { 1065 if (PrintOpto) { 1066 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1067 } 1068 return true; 1069 } else { 1070 return false; 1071 } 1072 } 1073 1074 static bool merge_point_safe(Node* region) { 1075 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1076 // having a PhiNode input. This sidesteps the dangerous case where the split 1077 // ConvI2LNode may become TOP if the input Value() does not 1078 // overlap the ConvI2L range, leaving a node which may not dominate its 1079 // uses. 1080 // A better fix for this problem can be found in the BugTraq entry, but 1081 // expediency for Mantis demands this hack. 1082 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1083 // split_if_with_blocks from splitting a block because we could not move around 1084 // the FastLockNode. 1085 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1086 Node* n = region->fast_out(i); 1087 if (n->is_Phi()) { 1088 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1089 Node* m = n->fast_out(j); 1090 if (m->is_FastLock()) 1091 return false; 1092 #ifdef _LP64 1093 if (m->Opcode() == Op_ConvI2L) 1094 return false; 1095 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1096 return false; 1097 } 1098 #endif 1099 } 1100 } 1101 } 1102 return true; 1103 } 1104 1105 1106 //------------------------------place_near_use--------------------------------- 1107 // Place some computation next to use but not inside inner loops. 1108 // For inner loop uses move it to the preheader area. 1109 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1110 IdealLoopTree *u_loop = get_loop( useblock ); 1111 if (u_loop->_irreducible) { 1112 return useblock; 1113 } 1114 if (u_loop->_child) { 1115 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1116 return u_loop->_head->in(LoopNode::EntryControl); 1117 } 1118 return useblock; 1119 } 1120 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1121 } 1122 1123 1124 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1125 if (!n->is_If() || n->is_CountedLoopEnd()) { 1126 return false; 1127 } 1128 if (!n->in(0)->is_Region()) { 1129 return false; 1130 } 1131 Node* region = n->in(0); 1132 Node* dom = idom(region); 1133 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1134 return false; 1135 } 1136 IfNode* dom_if = dom->as_If(); 1137 Node* proj_true = dom_if->proj_out(1); 1138 Node* proj_false = dom_if->proj_out(0); 1139 1140 for (uint i = 1; i < region->req(); i++) { 1141 if (is_dominator(proj_true, region->in(i))) { 1142 continue; 1143 } 1144 if (is_dominator(proj_false, region->in(i))) { 1145 continue; 1146 } 1147 return false; 1148 } 1149 1150 return true; 1151 } 1152 1153 1154 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1155 if (must_throttle_split_if()) { 1156 return false; 1157 } 1158 1159 // Do not do 'split-if' if irreducible loops are present. 1160 if (_has_irreducible_loops) { 1161 return false; 1162 } 1163 1164 if (merge_point_too_heavy(C, n_ctrl)) { 1165 return false; 1166 } 1167 1168 // Do not do 'split-if' if some paths are dead. First do dead code 1169 // elimination and then see if its still profitable. 1170 for (uint i = 1; i < n_ctrl->req(); i++) { 1171 if (n_ctrl->in(i) == C->top()) { 1172 return false; 1173 } 1174 } 1175 1176 // If trying to do a 'Split-If' at the loop head, it is only 1177 // profitable if the cmp folds up on BOTH paths. Otherwise we 1178 // risk peeling a loop forever. 1179 1180 // CNC - Disabled for now. Requires careful handling of loop 1181 // body selection for the cloned code. Also, make sure we check 1182 // for any input path not being in the same loop as n_ctrl. For 1183 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1184 // because the alternative loop entry points won't be converted 1185 // into LoopNodes. 1186 IdealLoopTree *n_loop = get_loop(n_ctrl); 1187 for (uint j = 1; j < n_ctrl->req(); j++) { 1188 if (get_loop(n_ctrl->in(j)) != n_loop) { 1189 return false; 1190 } 1191 } 1192 1193 // Check for safety of the merge point. 1194 if (!merge_point_safe(n_ctrl)) { 1195 return false; 1196 } 1197 1198 return true; 1199 } 1200 1201 // Detect if the node is the inner strip-mined loop 1202 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1203 static Node* is_inner_of_stripmined_loop(const Node* out) { 1204 Node* out_le = NULL; 1205 1206 if (out->is_CountedLoopEnd()) { 1207 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1208 1209 if (loop != NULL && loop->is_strip_mined()) { 1210 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1211 } 1212 } 1213 1214 return out_le; 1215 } 1216 1217 bool PhaseIdealLoop::flatten_array_element_type_check(Node *n) { 1218 // If the CmpP is a subtype check for a value that has just been 1219 // loaded from an array, the subtype checks guarantees the value 1220 // can't be stored in a flattened array and the load of the value 1221 // happens with a flattened array check then: push the type check 1222 // through the phi of the flattened array check. This needs special 1223 // logic because the subtype check's input is not a phi but a 1224 // LoadKlass that must first be cloned through the phi. 1225 if (n->Opcode() != Op_CmpP) { 1226 return false; 1227 } 1228 1229 Node* klassptr = n->in(1); 1230 Node* klasscon = n->in(2); 1231 1232 if (klassptr->is_DecodeNarrowPtr()) { 1233 klassptr = klassptr->in(1); 1234 } 1235 1236 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1237 return false; 1238 } 1239 1240 if (!klasscon->is_Con()) { 1241 return false; 1242 } 1243 1244 Node* addr = klassptr->in(MemNode::Address); 1245 1246 if (!addr->is_AddP()) { 1247 return false; 1248 } 1249 1250 intptr_t offset; 1251 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1252 1253 if (obj == NULL) { 1254 return false; 1255 } 1256 1257 assert(obj != NULL && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?"); 1258 if (obj->Opcode() == Op_CastPP) { 1259 obj = obj->in(1); 1260 } 1261 1262 if (!obj->is_Phi()) { 1263 return false; 1264 } 1265 1266 Node* region = obj->in(0); 1267 1268 Node* phi = PhiNode::make_blank(region, n->in(1)); 1269 for (uint i = 1; i < region->req(); i++) { 1270 Node* in = obj->in(i); 1271 Node* ctrl = get_ctrl(in); 1272 if (addr->in(AddPNode::Base) != obj) { 1273 Node* cast = addr->in(AddPNode::Base); 1274 assert(cast->Opcode() == Op_CastPP && cast->in(0) != NULL, "inconsistent subgraph"); 1275 Node* cast_clone = cast->clone(); 1276 cast_clone->set_req(0, region->in(i)); 1277 cast_clone->set_req(1, in); 1278 register_new_node(cast_clone, region->in(i)); 1279 _igvn.set_type(cast_clone, cast_clone->Value(&_igvn)); 1280 in = cast_clone; 1281 } 1282 Node* addr_clone = addr->clone(); 1283 addr_clone->set_req(AddPNode::Base, in); 1284 addr_clone->set_req(AddPNode::Address, in); 1285 register_new_node(addr_clone, ctrl); 1286 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1287 Node* klassptr_clone = klassptr->clone(); 1288 klassptr_clone->set_req(2, addr_clone); 1289 register_new_node(klassptr_clone, ctrl); 1290 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1291 if (klassptr != n->in(1)) { 1292 Node* decode = n->in(1); 1293 assert(decode->is_DecodeNarrowPtr(), "inconcistent subgraph"); 1294 Node* decode_clone = decode->clone(); 1295 decode_clone->set_req(1, klassptr_clone); 1296 register_new_node(decode_clone, ctrl); 1297 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1298 klassptr_clone = decode_clone; 1299 } 1300 phi->set_req(i, klassptr_clone); 1301 } 1302 register_new_node(phi, region); 1303 Node* orig = n->in(1); 1304 _igvn.replace_input_of(n, 1, phi); 1305 split_if_with_blocks_post(n); 1306 if (n->outcnt() != 0) { 1307 _igvn.replace_input_of(n, 1, orig); 1308 _igvn.remove_dead_node(phi); 1309 } 1310 return true; 1311 } 1312 1313 //------------------------------split_if_with_blocks_post---------------------- 1314 // Do the real work in a non-recursive function. CFG hackery wants to be 1315 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1316 // info. 1317 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1318 1319 if (flatten_array_element_type_check(n)) { 1320 return; 1321 } 1322 1323 // Cloning Cmp through Phi's involves the split-if transform. 1324 // FastLock is not used by an If 1325 if (n->is_Cmp() && !n->is_FastLock()) { 1326 Node *n_ctrl = get_ctrl(n); 1327 // Determine if the Node has inputs from some local Phi. 1328 // Returns the block to clone thru. 1329 Node *n_blk = has_local_phi_input(n); 1330 if (n_blk != n_ctrl) { 1331 return; 1332 } 1333 1334 if (!can_split_if(n_ctrl)) { 1335 return; 1336 } 1337 1338 if (n->outcnt() != 1) { 1339 return; // Multiple bool's from 1 compare? 1340 } 1341 Node *bol = n->unique_out(); 1342 assert(bol->is_Bool(), "expect a bool here"); 1343 if (bol->outcnt() != 1) { 1344 return;// Multiple branches from 1 compare? 1345 } 1346 Node *iff = bol->unique_out(); 1347 1348 // Check some safety conditions 1349 if (iff->is_If()) { // Classic split-if? 1350 if (iff->in(0) != n_ctrl) { 1351 return; // Compare must be in same blk as if 1352 } 1353 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1354 // Can't split CMove with different control edge. 1355 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1356 return; 1357 } 1358 if (get_ctrl(iff->in(2)) == n_ctrl || 1359 get_ctrl(iff->in(3)) == n_ctrl) { 1360 return; // Inputs not yet split-up 1361 } 1362 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1363 return; // Loop-invar test gates loop-varying CMOVE 1364 } 1365 } else { 1366 return; // some other kind of node, such as an Allocate 1367 } 1368 1369 // When is split-if profitable? Every 'win' on means some control flow 1370 // goes dead, so it's almost always a win. 1371 int policy = 0; 1372 // Split compare 'n' through the merge point if it is profitable 1373 Node *phi = split_thru_phi( n, n_ctrl, policy); 1374 if (!phi) { 1375 return; 1376 } 1377 1378 // Found a Phi to split thru! 1379 // Replace 'n' with the new phi 1380 _igvn.replace_node(n, phi); 1381 1382 // Now split the bool up thru the phi 1383 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1384 guarantee(bolphi != NULL, "null boolean phi node"); 1385 1386 _igvn.replace_node(bol, bolphi); 1387 assert(iff->in(1) == bolphi, ""); 1388 1389 if (bolphi->Value(&_igvn)->singleton()) { 1390 return; 1391 } 1392 1393 // Conditional-move? Must split up now 1394 if (!iff->is_If()) { 1395 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1396 _igvn.replace_node(iff, cmovphi); 1397 return; 1398 } 1399 1400 // Now split the IF 1401 do_split_if(iff); 1402 return; 1403 } 1404 1405 // Two identical ifs back to back can be merged 1406 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1407 Node *n_ctrl = n->in(0); 1408 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1409 IfNode* dom_if = idom(n_ctrl)->as_If(); 1410 Node* proj_true = dom_if->proj_out(1); 1411 Node* proj_false = dom_if->proj_out(0); 1412 Node* con_true = _igvn.makecon(TypeInt::ONE); 1413 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1414 1415 for (uint i = 1; i < n_ctrl->req(); i++) { 1416 if (is_dominator(proj_true, n_ctrl->in(i))) { 1417 bolphi->init_req(i, con_true); 1418 } else { 1419 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1420 bolphi->init_req(i, con_false); 1421 } 1422 } 1423 register_new_node(bolphi, n_ctrl); 1424 _igvn.replace_input_of(n, 1, bolphi); 1425 1426 // Now split the IF 1427 do_split_if(n); 1428 return; 1429 } 1430 1431 // Check for an IF ready to split; one that has its 1432 // condition codes input coming from a Phi at the block start. 1433 int n_op = n->Opcode(); 1434 1435 // Check for an IF being dominated by another IF same test 1436 if (n_op == Op_If || 1437 n_op == Op_RangeCheck) { 1438 Node *bol = n->in(1); 1439 uint max = bol->outcnt(); 1440 // Check for same test used more than once? 1441 if (max > 1 && bol->is_Bool()) { 1442 // Search up IDOMs to see if this IF is dominated. 1443 Node *cutoff = get_ctrl(bol); 1444 1445 // Now search up IDOMs till cutoff, looking for a dominating test 1446 Node *prevdom = n; 1447 Node *dom = idom(prevdom); 1448 while (dom != cutoff) { 1449 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1450 // It's invalid to move control dependent data nodes in the inner 1451 // strip-mined loop, because: 1452 // 1) break validation of LoopNode::verify_strip_mined() 1453 // 2) move code with side-effect in strip-mined loop 1454 // Move to the exit of outer strip-mined loop in that case. 1455 Node* out_le = is_inner_of_stripmined_loop(dom); 1456 if (out_le != NULL) { 1457 prevdom = out_le; 1458 } 1459 // Replace the dominated test with an obvious true or false. 1460 // Place it on the IGVN worklist for later cleanup. 1461 C->set_major_progress(); 1462 dominated_by(prevdom, n, false, true); 1463 #ifndef PRODUCT 1464 if( VerifyLoopOptimizations ) verify(); 1465 #endif 1466 return; 1467 } 1468 prevdom = dom; 1469 dom = idom(prevdom); 1470 } 1471 } 1472 } 1473 1474 // See if a shared loop-varying computation has no loop-varying uses. 1475 // Happens if something is only used for JVM state in uncommon trap exits, 1476 // like various versions of induction variable+offset. Clone the 1477 // computation per usage to allow it to sink out of the loop. 1478 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1479 Node *n_ctrl = get_ctrl(n); 1480 IdealLoopTree *n_loop = get_loop(n_ctrl); 1481 if( n_loop != _ltree_root ) { 1482 DUIterator_Fast imax, i = n->fast_outs(imax); 1483 for (; i < imax; i++) { 1484 Node* u = n->fast_out(i); 1485 if( !has_ctrl(u) ) break; // Found control user 1486 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1487 if( u_loop == n_loop ) break; // Found loop-varying use 1488 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1489 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1490 } 1491 bool did_break = (i < imax); // Did we break out of the previous loop? 1492 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1493 Node *late_load_ctrl = NULL; 1494 if (n->is_Load()) { 1495 // If n is a load, get and save the result from get_late_ctrl(), 1496 // to be later used in calculating the control for n's clones. 1497 clear_dom_lca_tags(); 1498 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1499 } 1500 // If n is a load, and the late control is the same as the current 1501 // control, then the cloning of n is a pointless exercise, because 1502 // GVN will ensure that we end up where we started. 1503 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1504 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1505 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1506 Node *u = n->last_out(j); // Clone private computation per use 1507 _igvn.rehash_node_delayed(u); 1508 Node *x = n->clone(); // Clone computation 1509 Node *x_ctrl = NULL; 1510 if( u->is_Phi() ) { 1511 // Replace all uses of normal nodes. Replace Phi uses 1512 // individually, so the separate Nodes can sink down 1513 // different paths. 1514 uint k = 1; 1515 while( u->in(k) != n ) k++; 1516 u->set_req( k, x ); 1517 // x goes next to Phi input path 1518 x_ctrl = u->in(0)->in(k); 1519 --j; 1520 } else { // Normal use 1521 // Replace all uses 1522 for( uint k = 0; k < u->req(); k++ ) { 1523 if( u->in(k) == n ) { 1524 u->set_req( k, x ); 1525 --j; 1526 } 1527 } 1528 x_ctrl = get_ctrl(u); 1529 } 1530 1531 // Find control for 'x' next to use but not inside inner loops. 1532 // For inner loop uses get the preheader area. 1533 x_ctrl = place_near_use(x_ctrl); 1534 1535 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) { 1536 continue; 1537 } 1538 1539 if (n->is_Load()) { 1540 // For loads, add a control edge to a CFG node outside of the loop 1541 // to force them to not combine and return back inside the loop 1542 // during GVN optimization (4641526). 1543 // 1544 // Because we are setting the actual control input, factor in 1545 // the result from get_late_ctrl() so we respect any 1546 // anti-dependences. (6233005). 1547 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1548 1549 // Don't allow the control input to be a CFG splitting node. 1550 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1551 // should only have IfTrueNode and IfFalseNode (4985384). 1552 x_ctrl = find_non_split_ctrl(x_ctrl); 1553 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1554 1555 x->set_req(0, x_ctrl); 1556 } 1557 register_new_node(x, x_ctrl); 1558 1559 // Some institutional knowledge is needed here: 'x' is 1560 // yanked because if the optimizer runs GVN on it all the 1561 // cloned x's will common up and undo this optimization and 1562 // be forced back in the loop. 1563 // I tried setting control edges on the x's to force them to 1564 // not combine, but the matching gets worried when it tries 1565 // to fold a StoreP and an AddP together (as part of an 1566 // address expression) and the AddP and StoreP have 1567 // different controls. 1568 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1569 } 1570 _igvn.remove_dead_node(n); 1571 } 1572 } 1573 } 1574 } 1575 1576 try_move_store_after_loop(n); 1577 1578 // Remove multiple allocations of the same value type 1579 if (n->is_ValueType()) { 1580 n->as_ValueType()->remove_redundant_allocations(&_igvn, this); 1581 return; // n is now dead 1582 } 1583 1584 // Check for Opaque2's who's loop has disappeared - who's input is in the 1585 // same loop nest as their output. Remove 'em, they are no longer useful. 1586 if( n_op == Op_Opaque2 && 1587 n->in(1) != NULL && 1588 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1589 _igvn.replace_node( n, n->in(1) ); 1590 } 1591 } 1592 1593 //------------------------------split_if_with_blocks--------------------------- 1594 // Check for aggressive application of 'split-if' optimization, 1595 // using basic block level info. 1596 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1597 Node* root = C->root(); 1598 visited.set(root->_idx); // first, mark root as visited 1599 // Do pre-visit work for root 1600 Node* n = split_if_with_blocks_pre(root); 1601 uint cnt = n->outcnt(); 1602 uint i = 0; 1603 1604 while (true) { 1605 // Visit all children 1606 if (i < cnt) { 1607 Node* use = n->raw_out(i); 1608 ++i; 1609 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1610 // Now do pre-visit work for this use 1611 use = split_if_with_blocks_pre(use); 1612 nstack.push(n, i); // Save parent and next use's index. 1613 n = use; // Process all children of current use. 1614 cnt = use->outcnt(); 1615 i = 0; 1616 } 1617 } 1618 else { 1619 // All of n's children have been processed, complete post-processing. 1620 if (cnt != 0 && !n->is_Con()) { 1621 assert(has_node(n), "no dead nodes"); 1622 split_if_with_blocks_post(n); 1623 } 1624 if (must_throttle_split_if()) { 1625 nstack.clear(); 1626 } 1627 if (nstack.is_empty()) { 1628 // Finished all nodes on stack. 1629 break; 1630 } 1631 // Get saved parent node and next use's index. Visit the rest of uses. 1632 n = nstack.node(); 1633 cnt = n->outcnt(); 1634 i = nstack.index(); 1635 nstack.pop(); 1636 } 1637 } 1638 } 1639 1640 1641 //============================================================================= 1642 // 1643 // C L O N E A L O O P B O D Y 1644 // 1645 1646 //------------------------------clone_iff-------------------------------------- 1647 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1648 // "Nearly" because all Nodes have been cloned from the original in the loop, 1649 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1650 // through the Phi recursively, and return a Bool. 1651 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1652 1653 // Convert this Phi into a Phi merging Bools 1654 uint i; 1655 for (i = 1; i < phi->req(); i++) { 1656 Node *b = phi->in(i); 1657 if (b->is_Phi()) { 1658 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1659 } else { 1660 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1661 } 1662 } 1663 1664 Node* n = phi->in(1); 1665 Node* sample_opaque = NULL; 1666 Node *sample_bool = NULL; 1667 if (n->Opcode() == Op_Opaque4) { 1668 sample_opaque = n; 1669 sample_bool = n->in(1); 1670 assert(sample_bool->is_Bool(), "wrong type"); 1671 } else { 1672 sample_bool = n; 1673 } 1674 Node *sample_cmp = sample_bool->in(1); 1675 1676 // Make Phis to merge the Cmp's inputs. 1677 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1678 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1679 for (i = 1; i < phi->req(); i++) { 1680 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1681 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1682 phi1->set_req(i, n1); 1683 phi2->set_req(i, n2); 1684 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1685 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1686 } 1687 // See if these Phis have been made before. 1688 // Register with optimizer 1689 Node *hit1 = _igvn.hash_find_insert(phi1); 1690 if (hit1) { // Hit, toss just made Phi 1691 _igvn.remove_dead_node(phi1); // Remove new phi 1692 assert(hit1->is_Phi(), "" ); 1693 phi1 = (PhiNode*)hit1; // Use existing phi 1694 } else { // Miss 1695 _igvn.register_new_node_with_optimizer(phi1); 1696 } 1697 Node *hit2 = _igvn.hash_find_insert(phi2); 1698 if (hit2) { // Hit, toss just made Phi 1699 _igvn.remove_dead_node(phi2); // Remove new phi 1700 assert(hit2->is_Phi(), "" ); 1701 phi2 = (PhiNode*)hit2; // Use existing phi 1702 } else { // Miss 1703 _igvn.register_new_node_with_optimizer(phi2); 1704 } 1705 // Register Phis with loop/block info 1706 set_ctrl(phi1, phi->in(0)); 1707 set_ctrl(phi2, phi->in(0)); 1708 // Make a new Cmp 1709 Node *cmp = sample_cmp->clone(); 1710 cmp->set_req(1, phi1); 1711 cmp->set_req(2, phi2); 1712 _igvn.register_new_node_with_optimizer(cmp); 1713 set_ctrl(cmp, phi->in(0)); 1714 1715 // Make a new Bool 1716 Node *b = sample_bool->clone(); 1717 b->set_req(1,cmp); 1718 _igvn.register_new_node_with_optimizer(b); 1719 set_ctrl(b, phi->in(0)); 1720 1721 if (sample_opaque != NULL) { 1722 Node* opaque = sample_opaque->clone(); 1723 opaque->set_req(1, b); 1724 _igvn.register_new_node_with_optimizer(opaque); 1725 set_ctrl(opaque, phi->in(0)); 1726 return opaque; 1727 } 1728 1729 assert(b->is_Bool(), ""); 1730 return b; 1731 } 1732 1733 //------------------------------clone_bool------------------------------------- 1734 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1735 // "Nearly" because all Nodes have been cloned from the original in the loop, 1736 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1737 // through the Phi recursively, and return a Bool. 1738 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1739 uint i; 1740 // Convert this Phi into a Phi merging Bools 1741 for( i = 1; i < phi->req(); i++ ) { 1742 Node *b = phi->in(i); 1743 if( b->is_Phi() ) { 1744 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1745 } else { 1746 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1747 } 1748 } 1749 1750 Node *sample_cmp = phi->in(1); 1751 1752 // Make Phis to merge the Cmp's inputs. 1753 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1754 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1755 for( uint j = 1; j < phi->req(); j++ ) { 1756 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1757 Node *n1, *n2; 1758 if( cmp_top->is_Cmp() ) { 1759 n1 = cmp_top->in(1); 1760 n2 = cmp_top->in(2); 1761 } else { 1762 n1 = n2 = cmp_top; 1763 } 1764 phi1->set_req( j, n1 ); 1765 phi2->set_req( j, n2 ); 1766 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1767 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1768 } 1769 1770 // See if these Phis have been made before. 1771 // Register with optimizer 1772 Node *hit1 = _igvn.hash_find_insert(phi1); 1773 if( hit1 ) { // Hit, toss just made Phi 1774 _igvn.remove_dead_node(phi1); // Remove new phi 1775 assert( hit1->is_Phi(), "" ); 1776 phi1 = (PhiNode*)hit1; // Use existing phi 1777 } else { // Miss 1778 _igvn.register_new_node_with_optimizer(phi1); 1779 } 1780 Node *hit2 = _igvn.hash_find_insert(phi2); 1781 if( hit2 ) { // Hit, toss just made Phi 1782 _igvn.remove_dead_node(phi2); // Remove new phi 1783 assert( hit2->is_Phi(), "" ); 1784 phi2 = (PhiNode*)hit2; // Use existing phi 1785 } else { // Miss 1786 _igvn.register_new_node_with_optimizer(phi2); 1787 } 1788 // Register Phis with loop/block info 1789 set_ctrl(phi1, phi->in(0)); 1790 set_ctrl(phi2, phi->in(0)); 1791 // Make a new Cmp 1792 Node *cmp = sample_cmp->clone(); 1793 cmp->set_req( 1, phi1 ); 1794 cmp->set_req( 2, phi2 ); 1795 _igvn.register_new_node_with_optimizer(cmp); 1796 set_ctrl(cmp, phi->in(0)); 1797 1798 assert( cmp->is_Cmp(), "" ); 1799 return (CmpNode*)cmp; 1800 } 1801 1802 //------------------------------sink_use--------------------------------------- 1803 // If 'use' was in the loop-exit block, it now needs to be sunk 1804 // below the post-loop merge point. 1805 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1806 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1807 set_ctrl(use, post_loop); 1808 for (DUIterator j = use->outs(); use->has_out(j); j++) 1809 sink_use(use->out(j), post_loop); 1810 } 1811 } 1812 1813 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1814 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1815 Node_List*& split_if_set, Node_List*& split_bool_set, 1816 Node_List*& split_cex_set, Node_List& worklist, 1817 uint new_counter, CloneLoopMode mode) { 1818 Node* nnn = old_new[old->_idx]; 1819 // Copy uses to a worklist, so I can munge the def-use info 1820 // with impunity. 1821 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1822 worklist.push(old->fast_out(j)); 1823 1824 while( worklist.size() ) { 1825 Node *use = worklist.pop(); 1826 if (!has_node(use)) continue; // Ignore dead nodes 1827 if (use->in(0) == C->top()) continue; 1828 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1829 // Check for data-use outside of loop - at least one of OLD or USE 1830 // must not be a CFG node. 1831 #ifdef ASSERT 1832 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1833 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1834 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1835 } 1836 #endif 1837 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1838 1839 // If the Data use is an IF, that means we have an IF outside of the 1840 // loop that is switching on a condition that is set inside of the 1841 // loop. Happens if people set a loop-exit flag; then test the flag 1842 // in the loop to break the loop, then test is again outside of the 1843 // loop to determine which way the loop exited. 1844 // Loop predicate If node connects to Bool node through Opaque1 node. 1845 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1846 // Since this code is highly unlikely, we lazily build the worklist 1847 // of such Nodes to go split. 1848 if (!split_if_set) { 1849 ResourceArea *area = Thread::current()->resource_area(); 1850 split_if_set = new Node_List(area); 1851 } 1852 split_if_set->push(use); 1853 } 1854 if (use->is_Bool()) { 1855 if (!split_bool_set) { 1856 ResourceArea *area = Thread::current()->resource_area(); 1857 split_bool_set = new Node_List(area); 1858 } 1859 split_bool_set->push(use); 1860 } 1861 if (use->Opcode() == Op_CreateEx) { 1862 if (!split_cex_set) { 1863 ResourceArea *area = Thread::current()->resource_area(); 1864 split_cex_set = new Node_List(area); 1865 } 1866 split_cex_set->push(use); 1867 } 1868 1869 1870 // Get "block" use is in 1871 uint idx = 0; 1872 while( use->in(idx) != old ) idx++; 1873 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1874 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1875 Node *cfg = prev->_idx >= new_counter 1876 ? prev->in(2) 1877 : idom(prev); 1878 if( use->is_Phi() ) // Phi use is in prior block 1879 cfg = prev->in(idx); // NOT in block of Phi itself 1880 if (cfg->is_top()) { // Use is dead? 1881 _igvn.replace_input_of(use, idx, C->top()); 1882 continue; 1883 } 1884 1885 // If use is referenced through control edge... (idx == 0) 1886 if (mode == IgnoreStripMined && idx == 0) { 1887 LoopNode *head = loop->_head->as_Loop(); 1888 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1889 // That node is outside the inner loop, leave it outside the 1890 // outer loop as well to not confuse verification code. 1891 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1892 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1893 continue; 1894 } 1895 } 1896 1897 while(!outer_loop->is_member(get_loop(cfg))) { 1898 prev = cfg; 1899 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1900 } 1901 // If the use occurs after merging several exits from the loop, then 1902 // old value must have dominated all those exits. Since the same old 1903 // value was used on all those exits we did not need a Phi at this 1904 // merge point. NOW we do need a Phi here. Each loop exit value 1905 // is now merged with the peeled body exit; each exit gets its own 1906 // private Phi and those Phis need to be merged here. 1907 Node *phi; 1908 if( prev->is_Region() ) { 1909 if( idx == 0 ) { // Updating control edge? 1910 phi = prev; // Just use existing control 1911 } else { // Else need a new Phi 1912 phi = PhiNode::make( prev, old ); 1913 // Now recursively fix up the new uses of old! 1914 for( uint i = 1; i < prev->req(); i++ ) { 1915 worklist.push(phi); // Onto worklist once for each 'old' input 1916 } 1917 } 1918 } else { 1919 // Get new RegionNode merging old and new loop exits 1920 prev = old_new[prev->_idx]; 1921 assert( prev, "just made this in step 7" ); 1922 if( idx == 0) { // Updating control edge? 1923 phi = prev; // Just use existing control 1924 } else { // Else need a new Phi 1925 // Make a new Phi merging data values properly 1926 phi = PhiNode::make( prev, old ); 1927 phi->set_req( 1, nnn ); 1928 } 1929 } 1930 // If inserting a new Phi, check for prior hits 1931 if( idx != 0 ) { 1932 Node *hit = _igvn.hash_find_insert(phi); 1933 if( hit == NULL ) { 1934 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1935 } else { // or 1936 // Remove the new phi from the graph and use the hit 1937 _igvn.remove_dead_node(phi); 1938 phi = hit; // Use existing phi 1939 } 1940 set_ctrl(phi, prev); 1941 } 1942 // Make 'use' use the Phi instead of the old loop body exit value 1943 _igvn.replace_input_of(use, idx, phi); 1944 if( use->_idx >= new_counter ) { // If updating new phis 1945 // Not needed for correctness, but prevents a weak assert 1946 // in AddPNode from tripping (when we end up with different 1947 // base & derived Phis that will become the same after 1948 // IGVN does CSE). 1949 Node *hit = _igvn.hash_find_insert(use); 1950 if( hit ) // Go ahead and re-hash for hits. 1951 _igvn.replace_node( use, hit ); 1952 } 1953 1954 // If 'use' was in the loop-exit block, it now needs to be sunk 1955 // below the post-loop merge point. 1956 sink_use( use, prev ); 1957 } 1958 } 1959 } 1960 1961 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1962 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1963 bool check_old_new) { 1964 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1965 Node* u = n->fast_out(j); 1966 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1967 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1968 Node* c = phase->get_ctrl(u); 1969 IdealLoopTree* u_loop = phase->get_loop(c); 1970 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1971 if (outer_loop->is_member(u_loop)) { 1972 wq.push(u); 1973 } 1974 } 1975 } 1976 } 1977 1978 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1979 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1980 Node_List& extra_data_nodes) { 1981 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1982 CountedLoopNode* cl = head->as_CountedLoop(); 1983 Node* l = cl->outer_loop(); 1984 Node* tail = cl->outer_loop_tail(); 1985 IfNode* le = cl->outer_loop_end(); 1986 Node* sfpt = cl->outer_safepoint(); 1987 CountedLoopEndNode* cle = cl->loopexit(); 1988 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1989 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1990 Node* cle_out = cle->proj_out(false); 1991 1992 Node* new_sfpt = NULL; 1993 Node* new_cle_out = cle_out->clone(); 1994 old_new.map(cle_out->_idx, new_cle_out); 1995 if (mode == CloneIncludesStripMined) { 1996 // clone outer loop body 1997 Node* new_l = l->clone(); 1998 Node* new_tail = tail->clone(); 1999 IfNode* new_le = le->clone()->as_If(); 2000 new_sfpt = sfpt->clone(); 2001 2002 set_loop(new_l, outer_loop->_parent); 2003 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2004 set_loop(new_cle_out, outer_loop->_parent); 2005 set_idom(new_cle_out, new_cle, dd); 2006 set_loop(new_sfpt, outer_loop->_parent); 2007 set_idom(new_sfpt, new_cle_out, dd); 2008 set_loop(new_le, outer_loop->_parent); 2009 set_idom(new_le, new_sfpt, dd); 2010 set_loop(new_tail, outer_loop->_parent); 2011 set_idom(new_tail, new_le, dd); 2012 set_idom(new_cl, new_l, dd); 2013 2014 old_new.map(l->_idx, new_l); 2015 old_new.map(tail->_idx, new_tail); 2016 old_new.map(le->_idx, new_le); 2017 old_new.map(sfpt->_idx, new_sfpt); 2018 2019 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2020 new_l->set_req(0, new_l); 2021 new_tail->set_req(0, new_le); 2022 new_le->set_req(0, new_sfpt); 2023 new_sfpt->set_req(0, new_cle_out); 2024 new_cle_out->set_req(0, new_cle); 2025 new_cl->set_req(LoopNode::EntryControl, new_l); 2026 2027 _igvn.register_new_node_with_optimizer(new_l); 2028 _igvn.register_new_node_with_optimizer(new_tail); 2029 _igvn.register_new_node_with_optimizer(new_le); 2030 } else { 2031 Node *newhead = old_new[loop->_head->_idx]; 2032 newhead->as_Loop()->clear_strip_mined(); 2033 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2034 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2035 } 2036 // Look at data node that were assigned a control in the outer 2037 // loop: they are kept in the outer loop by the safepoint so start 2038 // from the safepoint node's inputs. 2039 IdealLoopTree* outer_loop = get_loop(l); 2040 Node_Stack stack(2); 2041 stack.push(sfpt, 1); 2042 uint new_counter = C->unique(); 2043 while (stack.size() > 0) { 2044 Node* n = stack.node(); 2045 uint i = stack.index(); 2046 while (i < n->req() && 2047 (n->in(i) == NULL || 2048 !has_ctrl(n->in(i)) || 2049 get_loop(get_ctrl(n->in(i))) != outer_loop || 2050 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2051 i++; 2052 } 2053 if (i < n->req()) { 2054 stack.set_index(i+1); 2055 stack.push(n->in(i), 0); 2056 } else { 2057 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2058 Node* m = n == sfpt ? new_sfpt : n->clone(); 2059 if (m != NULL) { 2060 for (uint i = 0; i < n->req(); i++) { 2061 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 2062 m->set_req(i, old_new[m->in(i)->_idx]); 2063 } 2064 } 2065 } else { 2066 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2067 } 2068 if (n != sfpt) { 2069 extra_data_nodes.push(n); 2070 _igvn.register_new_node_with_optimizer(m); 2071 assert(get_ctrl(n) == cle_out, "what other control?"); 2072 set_ctrl(m, new_cle_out); 2073 old_new.map(n->_idx, m); 2074 } 2075 stack.pop(); 2076 } 2077 } 2078 if (mode == CloneIncludesStripMined) { 2079 _igvn.register_new_node_with_optimizer(new_sfpt); 2080 _igvn.register_new_node_with_optimizer(new_cle_out); 2081 } 2082 // Some other transformation may have pessimistically assign some 2083 // data nodes to the outer loop. Set their control so they are out 2084 // of the outer loop. 2085 ResourceMark rm; 2086 Unique_Node_List wq; 2087 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2088 Node* old = extra_data_nodes.at(i); 2089 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 2090 } 2091 Node* new_ctrl = cl->outer_loop_exit(); 2092 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2093 for (uint i = 0; i < wq.size(); i++) { 2094 Node* n = wq.at(i); 2095 set_ctrl(n, new_ctrl); 2096 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 2097 } 2098 } else { 2099 Node *newhead = old_new[loop->_head->_idx]; 2100 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2101 } 2102 } 2103 2104 //------------------------------clone_loop------------------------------------- 2105 // 2106 // C L O N E A L O O P B O D Y 2107 // 2108 // This is the basic building block of the loop optimizations. It clones an 2109 // entire loop body. It makes an old_new loop body mapping; with this mapping 2110 // you can find the new-loop equivalent to an old-loop node. All new-loop 2111 // nodes are exactly equal to their old-loop counterparts, all edges are the 2112 // same. All exits from the old-loop now have a RegionNode that merges the 2113 // equivalent new-loop path. This is true even for the normal "loop-exit" 2114 // condition. All uses of loop-invariant old-loop values now come from (one 2115 // or more) Phis that merge their new-loop equivalents. 2116 // 2117 // This operation leaves the graph in an illegal state: there are two valid 2118 // control edges coming from the loop pre-header to both loop bodies. I'll 2119 // definitely have to hack the graph after running this transform. 2120 // 2121 // From this building block I will further edit edges to perform loop peeling 2122 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2123 // 2124 // Parameter side_by_size_idom: 2125 // When side_by_size_idom is NULL, the dominator tree is constructed for 2126 // the clone loop to dominate the original. Used in construction of 2127 // pre-main-post loop sequence. 2128 // When nonnull, the clone and original are side-by-side, both are 2129 // dominated by the side_by_side_idom node. Used in construction of 2130 // unswitched loops. 2131 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2132 CloneLoopMode mode, Node* side_by_side_idom) { 2133 2134 LoopNode* head = loop->_head->as_Loop(); 2135 head->verify_strip_mined(1); 2136 2137 if (C->do_vector_loop() && PrintOpto) { 2138 const char* mname = C->method()->name()->as_quoted_ascii(); 2139 if (mname != NULL) { 2140 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2141 } 2142 } 2143 2144 CloneMap& cm = C->clone_map(); 2145 Dict* dict = cm.dict(); 2146 if (C->do_vector_loop()) { 2147 cm.set_clone_idx(cm.max_gen()+1); 2148 #ifndef PRODUCT 2149 if (PrintOpto) { 2150 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2151 loop->dump_head(); 2152 } 2153 #endif 2154 } 2155 2156 // Step 1: Clone the loop body. Make the old->new mapping. 2157 uint i; 2158 for( i = 0; i < loop->_body.size(); i++ ) { 2159 Node *old = loop->_body.at(i); 2160 Node *nnn = old->clone(); 2161 old_new.map( old->_idx, nnn ); 2162 if (C->do_vector_loop()) { 2163 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2164 } 2165 _igvn.register_new_node_with_optimizer(nnn); 2166 } 2167 2168 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2169 2170 // Step 2: Fix the edges in the new body. If the old input is outside the 2171 // loop use it. If the old input is INside the loop, use the corresponding 2172 // new node instead. 2173 for( i = 0; i < loop->_body.size(); i++ ) { 2174 Node *old = loop->_body.at(i); 2175 Node *nnn = old_new[old->_idx]; 2176 // Fix CFG/Loop controlling the new node 2177 if (has_ctrl(old)) { 2178 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2179 } else { 2180 set_loop(nnn, outer_loop->_parent); 2181 if (old->outcnt() > 0) { 2182 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2183 } 2184 } 2185 // Correct edges to the new node 2186 for( uint j = 0; j < nnn->req(); j++ ) { 2187 Node *n = nnn->in(j); 2188 if( n ) { 2189 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2190 if( loop->is_member( old_in_loop ) ) 2191 nnn->set_req(j, old_new[n->_idx]); 2192 } 2193 } 2194 _igvn.hash_find_insert(nnn); 2195 } 2196 2197 ResourceArea *area = Thread::current()->resource_area(); 2198 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop 2199 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2200 2201 // Step 3: Now fix control uses. Loop varying control uses have already 2202 // been fixed up (as part of all input edges in Step 2). Loop invariant 2203 // control uses must be either an IfFalse or an IfTrue. Make a merge 2204 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2205 // refer to this. 2206 Node_List worklist(area); 2207 uint new_counter = C->unique(); 2208 for( i = 0; i < loop->_body.size(); i++ ) { 2209 Node* old = loop->_body.at(i); 2210 if( !old->is_CFG() ) continue; 2211 2212 // Copy uses to a worklist, so I can munge the def-use info 2213 // with impunity. 2214 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2215 worklist.push(old->fast_out(j)); 2216 2217 while( worklist.size() ) { // Visit all uses 2218 Node *use = worklist.pop(); 2219 if (!has_node(use)) continue; // Ignore dead nodes 2220 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2221 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2222 // Both OLD and USE are CFG nodes here. 2223 assert( use->is_Proj(), "" ); 2224 Node* nnn = old_new[old->_idx]; 2225 2226 Node* newuse = NULL; 2227 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2228 CountedLoopNode* cl = head->as_CountedLoop(); 2229 CountedLoopEndNode* cle = cl->loopexit(); 2230 Node* cle_out = cle->proj_out_or_null(false); 2231 if (use == cle_out) { 2232 IfNode* le = cl->outer_loop_end(); 2233 use = le->proj_out(false); 2234 use_loop = get_loop(use); 2235 if (mode == CloneIncludesStripMined) { 2236 nnn = old_new[le->_idx]; 2237 } else { 2238 newuse = old_new[cle_out->_idx]; 2239 } 2240 } 2241 } 2242 if (newuse == NULL) { 2243 newuse = use->clone(); 2244 } 2245 2246 // Clone the loop exit control projection 2247 if (C->do_vector_loop()) { 2248 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2249 } 2250 newuse->set_req(0,nnn); 2251 _igvn.register_new_node_with_optimizer(newuse); 2252 set_loop(newuse, use_loop); 2253 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2254 2255 // We need a Region to merge the exit from the peeled body and the 2256 // exit from the old loop body. 2257 RegionNode *r = new RegionNode(3); 2258 // Map the old use to the new merge point 2259 old_new.map( use->_idx, r ); 2260 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2261 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2262 2263 // The original user of 'use' uses 'r' instead. 2264 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2265 Node* useuse = use->last_out(l); 2266 _igvn.rehash_node_delayed(useuse); 2267 uint uses_found = 0; 2268 if( useuse->in(0) == use ) { 2269 useuse->set_req(0, r); 2270 uses_found++; 2271 if( useuse->is_CFG() ) { 2272 assert( dom_depth(useuse) > dd_r, "" ); 2273 set_idom(useuse, r, dom_depth(useuse)); 2274 } 2275 } 2276 for( uint k = 1; k < useuse->req(); k++ ) { 2277 if( useuse->in(k) == use ) { 2278 useuse->set_req(k, r); 2279 uses_found++; 2280 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2281 assert(dom_depth(useuse) > dd_r , ""); 2282 set_idom(useuse, r, dom_depth(useuse)); 2283 } 2284 } 2285 } 2286 l -= uses_found; // we deleted 1 or more copies of this edge 2287 } 2288 2289 // Now finish up 'r' 2290 r->set_req( 1, newuse ); 2291 r->set_req( 2, use ); 2292 _igvn.register_new_node_with_optimizer(r); 2293 set_loop(r, use_loop); 2294 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2295 } // End of if a loop-exit test 2296 } 2297 } 2298 2299 // Step 4: If loop-invariant use is not control, it must be dominated by a 2300 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2301 // there if needed. Make a Phi there merging old and new used values. 2302 Node_List *split_if_set = NULL; 2303 Node_List *split_bool_set = NULL; 2304 Node_List *split_cex_set = NULL; 2305 for( i = 0; i < loop->_body.size(); i++ ) { 2306 Node* old = loop->_body.at(i); 2307 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2308 split_bool_set, split_cex_set, worklist, new_counter, 2309 mode); 2310 } 2311 2312 for (i = 0; i < extra_data_nodes.size(); i++) { 2313 Node* old = extra_data_nodes.at(i); 2314 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2315 split_bool_set, split_cex_set, worklist, new_counter, 2316 mode); 2317 } 2318 2319 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2320 // the loop uses a condition set in the loop. The original IF probably 2321 // takes control from one or more OLD Regions (which in turn get from NEW 2322 // Regions). In any case, there will be a set of Phis for each merge point 2323 // from the IF up to where the original BOOL def exists the loop. 2324 if (split_if_set) { 2325 while (split_if_set->size()) { 2326 Node *iff = split_if_set->pop(); 2327 if (iff->in(1)->is_Phi()) { 2328 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2329 _igvn.replace_input_of(iff, 1, b); 2330 } 2331 } 2332 } 2333 if (split_bool_set) { 2334 while (split_bool_set->size()) { 2335 Node *b = split_bool_set->pop(); 2336 Node *phi = b->in(1); 2337 assert(phi->is_Phi(), ""); 2338 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2339 _igvn.replace_input_of(b, 1, cmp); 2340 } 2341 } 2342 if (split_cex_set) { 2343 while (split_cex_set->size()) { 2344 Node *b = split_cex_set->pop(); 2345 assert(b->in(0)->is_Region(), ""); 2346 assert(b->in(1)->is_Phi(), ""); 2347 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2348 split_up(b, b->in(0), NULL); 2349 } 2350 } 2351 2352 } 2353 2354 2355 //---------------------- stride_of_possible_iv ------------------------------------- 2356 // Looks for an iff/bool/comp with one operand of the compare 2357 // being a cycle involving an add and a phi, 2358 // with an optional truncation (left-shift followed by a right-shift) 2359 // of the add. Returns zero if not an iv. 2360 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2361 Node* trunc1 = NULL; 2362 Node* trunc2 = NULL; 2363 const TypeInt* ttype = NULL; 2364 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2365 return 0; 2366 } 2367 BoolNode* bl = iff->in(1)->as_Bool(); 2368 Node* cmp = bl->in(1); 2369 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2370 return 0; 2371 } 2372 // Must have an invariant operand 2373 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2374 return 0; 2375 } 2376 Node* add2 = NULL; 2377 Node* cmp1 = cmp->in(1); 2378 if (cmp1->is_Phi()) { 2379 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2380 Node* phi = cmp1; 2381 for (uint i = 1; i < phi->req(); i++) { 2382 Node* in = phi->in(i); 2383 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2384 &trunc1, &trunc2, &ttype); 2385 if (add && add->in(1) == phi) { 2386 add2 = add->in(2); 2387 break; 2388 } 2389 } 2390 } else { 2391 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2392 Node* addtrunc = cmp1; 2393 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2394 &trunc1, &trunc2, &ttype); 2395 if (add && add->in(1)->is_Phi()) { 2396 Node* phi = add->in(1); 2397 for (uint i = 1; i < phi->req(); i++) { 2398 if (phi->in(i) == addtrunc) { 2399 add2 = add->in(2); 2400 break; 2401 } 2402 } 2403 } 2404 } 2405 if (add2 != NULL) { 2406 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2407 if (add2t->is_con()) { 2408 return add2t->get_con(); 2409 } 2410 } 2411 return 0; 2412 } 2413 2414 2415 //---------------------- stay_in_loop ------------------------------------- 2416 // Return the (unique) control output node that's in the loop (if it exists.) 2417 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2418 Node* unique = NULL; 2419 if (!n) return NULL; 2420 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2421 Node* use = n->fast_out(i); 2422 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2423 if (unique != NULL) { 2424 return NULL; 2425 } 2426 unique = use; 2427 } 2428 } 2429 return unique; 2430 } 2431 2432 //------------------------------ register_node ------------------------------------- 2433 // Utility to register node "n" with PhaseIdealLoop 2434 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2435 _igvn.register_new_node_with_optimizer(n); 2436 loop->_body.push(n); 2437 if (n->is_CFG()) { 2438 set_loop(n, loop); 2439 set_idom(n, pred, ddepth); 2440 } else { 2441 set_ctrl(n, pred); 2442 } 2443 } 2444 2445 //------------------------------ proj_clone ------------------------------------- 2446 // Utility to create an if-projection 2447 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2448 ProjNode* c = p->clone()->as_Proj(); 2449 c->set_req(0, iff); 2450 return c; 2451 } 2452 2453 //------------------------------ short_circuit_if ------------------------------------- 2454 // Force the iff control output to be the live_proj 2455 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2456 guarantee(live_proj != NULL, "null projection"); 2457 int proj_con = live_proj->_con; 2458 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2459 Node *con = _igvn.intcon(proj_con); 2460 set_ctrl(con, C->root()); 2461 if (iff) { 2462 iff->set_req(1, con); 2463 } 2464 return con; 2465 } 2466 2467 //------------------------------ insert_if_before_proj ------------------------------------- 2468 // Insert a new if before an if projection (* - new node) 2469 // 2470 // before 2471 // if(test) 2472 // / \ 2473 // v v 2474 // other-proj proj (arg) 2475 // 2476 // after 2477 // if(test) 2478 // / \ 2479 // / v 2480 // | * proj-clone 2481 // v | 2482 // other-proj v 2483 // * new_if(relop(cmp[IU](left,right))) 2484 // / \ 2485 // v v 2486 // * new-proj proj 2487 // (returned) 2488 // 2489 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2490 IfNode* iff = proj->in(0)->as_If(); 2491 IdealLoopTree *loop = get_loop(proj); 2492 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2493 int ddepth = dom_depth(proj); 2494 2495 _igvn.rehash_node_delayed(iff); 2496 _igvn.rehash_node_delayed(proj); 2497 2498 proj->set_req(0, NULL); // temporary disconnect 2499 ProjNode* proj2 = proj_clone(proj, iff); 2500 register_node(proj2, loop, iff, ddepth); 2501 2502 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2503 register_node(cmp, loop, proj2, ddepth); 2504 2505 BoolNode* bol = new BoolNode(cmp, relop); 2506 register_node(bol, loop, proj2, ddepth); 2507 2508 int opcode = iff->Opcode(); 2509 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2510 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2511 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2512 register_node(new_if, loop, proj2, ddepth); 2513 2514 proj->set_req(0, new_if); // reattach 2515 set_idom(proj, new_if, ddepth); 2516 2517 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2518 guarantee(new_exit != NULL, "null exit node"); 2519 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2520 2521 return new_exit; 2522 } 2523 2524 //------------------------------ insert_region_before_proj ------------------------------------- 2525 // Insert a region before an if projection (* - new node) 2526 // 2527 // before 2528 // if(test) 2529 // / | 2530 // v | 2531 // proj v 2532 // other-proj 2533 // 2534 // after 2535 // if(test) 2536 // / | 2537 // v | 2538 // * proj-clone v 2539 // | other-proj 2540 // v 2541 // * new-region 2542 // | 2543 // v 2544 // * dum_if 2545 // / \ 2546 // v \ 2547 // * dum-proj v 2548 // proj 2549 // 2550 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2551 IfNode* iff = proj->in(0)->as_If(); 2552 IdealLoopTree *loop = get_loop(proj); 2553 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2554 int ddepth = dom_depth(proj); 2555 2556 _igvn.rehash_node_delayed(iff); 2557 _igvn.rehash_node_delayed(proj); 2558 2559 proj->set_req(0, NULL); // temporary disconnect 2560 ProjNode* proj2 = proj_clone(proj, iff); 2561 register_node(proj2, loop, iff, ddepth); 2562 2563 RegionNode* reg = new RegionNode(2); 2564 reg->set_req(1, proj2); 2565 register_node(reg, loop, iff, ddepth); 2566 2567 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2568 register_node(dum_if, loop, reg, ddepth); 2569 2570 proj->set_req(0, dum_if); // reattach 2571 set_idom(proj, dum_if, ddepth); 2572 2573 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2574 register_node(dum_proj, loop, dum_if, ddepth); 2575 2576 return reg; 2577 } 2578 2579 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2580 // Clone a signed compare loop exit from an unsigned compare and 2581 // insert it before the unsigned cmp on the stay-in-loop path. 2582 // All new nodes inserted in the dominator tree between the original 2583 // if and it's projections. The original if test is replaced with 2584 // a constant to force the stay-in-loop path. 2585 // 2586 // This is done to make sure that the original if and it's projections 2587 // still dominate the same set of control nodes, that the ctrl() relation 2588 // from data nodes to them is preserved, and that their loop nesting is 2589 // preserved. 2590 // 2591 // before 2592 // if(i <u limit) unsigned compare loop exit 2593 // / | 2594 // v v 2595 // exit-proj stay-in-loop-proj 2596 // 2597 // after 2598 // if(stay-in-loop-const) original if 2599 // / | 2600 // / v 2601 // / if(i < limit) new signed test 2602 // / / | 2603 // / / v 2604 // / / if(i <u limit) new cloned unsigned test 2605 // / / / | 2606 // v v v | 2607 // region | 2608 // | | 2609 // dum-if | 2610 // / | | 2611 // ether | | 2612 // v v 2613 // exit-proj stay-in-loop-proj 2614 // 2615 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2616 const bool Signed = true; 2617 const bool Unsigned = false; 2618 2619 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2620 if (bol->_test._test != BoolTest::lt) return NULL; 2621 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2622 if (cmpu->Opcode() != Op_CmpU) return NULL; 2623 int stride = stride_of_possible_iv(if_cmpu); 2624 if (stride == 0) return NULL; 2625 2626 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2627 guarantee(lp_proj != NULL, "null loop node"); 2628 2629 ProjNode* lp_continue = lp_proj->as_Proj(); 2630 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2631 2632 Node* limit = NULL; 2633 if (stride > 0) { 2634 limit = cmpu->in(2); 2635 } else { 2636 limit = _igvn.makecon(TypeInt::ZERO); 2637 set_ctrl(limit, C->root()); 2638 } 2639 // Create a new region on the exit path 2640 RegionNode* reg = insert_region_before_proj(lp_exit); 2641 guarantee(reg != NULL, "null region node"); 2642 2643 // Clone the if-cmpu-true-false using a signed compare 2644 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2645 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2646 reg->add_req(cmpi_exit); 2647 2648 // Clone the if-cmpu-true-false 2649 BoolTest::mask rel_u = bol->_test._test; 2650 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2651 reg->add_req(cmpu_exit); 2652 2653 // Force original if to stay in loop. 2654 short_circuit_if(if_cmpu, lp_continue); 2655 2656 return cmpi_exit->in(0)->as_If(); 2657 } 2658 2659 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2660 // Remove a previously inserted signed compare loop exit. 2661 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2662 Node* lp_proj = stay_in_loop(if_cmp, loop); 2663 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2664 stay_in_loop(lp_proj, loop)->is_If() && 2665 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2666 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2667 set_ctrl(con, C->root()); 2668 if_cmp->set_req(1, con); 2669 } 2670 2671 //------------------------------ scheduled_nodelist ------------------------------------- 2672 // Create a post order schedule of nodes that are in the 2673 // "member" set. The list is returned in "sched". 2674 // The first node in "sched" is the loop head, followed by 2675 // nodes which have no inputs in the "member" set, and then 2676 // followed by the nodes that have an immediate input dependence 2677 // on a node in "sched". 2678 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2679 2680 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2681 Arena *a = Thread::current()->resource_area(); 2682 VectorSet visited(a); 2683 Node_Stack nstack(a, loop->_body.size()); 2684 2685 Node* n = loop->_head; // top of stack is cached in "n" 2686 uint idx = 0; 2687 visited.set(n->_idx); 2688 2689 // Initially push all with no inputs from within member set 2690 for(uint i = 0; i < loop->_body.size(); i++ ) { 2691 Node *elt = loop->_body.at(i); 2692 if (member.test(elt->_idx)) { 2693 bool found = false; 2694 for (uint j = 0; j < elt->req(); j++) { 2695 Node* def = elt->in(j); 2696 if (def && member.test(def->_idx) && def != elt) { 2697 found = true; 2698 break; 2699 } 2700 } 2701 if (!found && elt != loop->_head) { 2702 nstack.push(n, idx); 2703 n = elt; 2704 assert(!visited.test(n->_idx), "not seen yet"); 2705 visited.set(n->_idx); 2706 } 2707 } 2708 } 2709 2710 // traverse out's that are in the member set 2711 while (true) { 2712 if (idx < n->outcnt()) { 2713 Node* use = n->raw_out(idx); 2714 idx++; 2715 if (!visited.test_set(use->_idx)) { 2716 if (member.test(use->_idx)) { 2717 nstack.push(n, idx); 2718 n = use; 2719 idx = 0; 2720 } 2721 } 2722 } else { 2723 // All outputs processed 2724 sched.push(n); 2725 if (nstack.is_empty()) break; 2726 n = nstack.node(); 2727 idx = nstack.index(); 2728 nstack.pop(); 2729 } 2730 } 2731 } 2732 2733 2734 //------------------------------ has_use_in_set ------------------------------------- 2735 // Has a use in the vector set 2736 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2737 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2738 Node* use = n->fast_out(j); 2739 if (vset.test(use->_idx)) { 2740 return true; 2741 } 2742 } 2743 return false; 2744 } 2745 2746 2747 //------------------------------ has_use_internal_to_set ------------------------------------- 2748 // Has use internal to the vector set (ie. not in a phi at the loop head) 2749 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2750 Node* head = loop->_head; 2751 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2752 Node* use = n->fast_out(j); 2753 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2754 return true; 2755 } 2756 } 2757 return false; 2758 } 2759 2760 2761 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2762 // clone "n" for uses that are outside of loop 2763 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2764 int cloned = 0; 2765 assert(worklist.size() == 0, "should be empty"); 2766 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2767 Node* use = n->fast_out(j); 2768 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2769 worklist.push(use); 2770 } 2771 } 2772 while( worklist.size() ) { 2773 Node *use = worklist.pop(); 2774 if (!has_node(use) || use->in(0) == C->top()) continue; 2775 uint j; 2776 for (j = 0; j < use->req(); j++) { 2777 if (use->in(j) == n) break; 2778 } 2779 assert(j < use->req(), "must be there"); 2780 2781 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2782 Node* n_clone = n->clone(); 2783 _igvn.replace_input_of(use, j, n_clone); 2784 cloned++; 2785 Node* use_c; 2786 if (!use->is_Phi()) { 2787 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2788 } else { 2789 // Use in a phi is considered a use in the associated predecessor block 2790 use_c = use->in(0)->in(j); 2791 } 2792 set_ctrl(n_clone, use_c); 2793 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2794 get_loop(use_c)->_body.push(n_clone); 2795 _igvn.register_new_node_with_optimizer(n_clone); 2796 #ifndef PRODUCT 2797 if (TracePartialPeeling) { 2798 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2799 } 2800 #endif 2801 } 2802 return cloned; 2803 } 2804 2805 2806 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2807 // clone "n" for special uses that are in the not_peeled region. 2808 // If these def-uses occur in separate blocks, the code generator 2809 // marks the method as not compilable. For example, if a "BoolNode" 2810 // is in a different basic block than the "IfNode" that uses it, then 2811 // the compilation is aborted in the code generator. 2812 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2813 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2814 if (n->is_Phi() || n->is_Load()) { 2815 return; 2816 } 2817 assert(worklist.size() == 0, "should be empty"); 2818 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2819 Node* use = n->fast_out(j); 2820 if ( not_peel.test(use->_idx) && 2821 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2822 use->in(1) == n) { 2823 worklist.push(use); 2824 } 2825 } 2826 if (worklist.size() > 0) { 2827 // clone "n" and insert it between inputs of "n" and the use 2828 Node* n_clone = n->clone(); 2829 loop->_body.push(n_clone); 2830 _igvn.register_new_node_with_optimizer(n_clone); 2831 set_ctrl(n_clone, get_ctrl(n)); 2832 sink_list.push(n_clone); 2833 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2834 #ifndef PRODUCT 2835 if (TracePartialPeeling) { 2836 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2837 } 2838 #endif 2839 while( worklist.size() ) { 2840 Node *use = worklist.pop(); 2841 _igvn.rehash_node_delayed(use); 2842 for (uint j = 1; j < use->req(); j++) { 2843 if (use->in(j) == n) { 2844 use->set_req(j, n_clone); 2845 } 2846 } 2847 } 2848 } 2849 } 2850 2851 2852 //------------------------------ insert_phi_for_loop ------------------------------------- 2853 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2854 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2855 Node *phi = PhiNode::make(lp, back_edge_val); 2856 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2857 // Use existing phi if it already exists 2858 Node *hit = _igvn.hash_find_insert(phi); 2859 if( hit == NULL ) { 2860 _igvn.register_new_node_with_optimizer(phi); 2861 set_ctrl(phi, lp); 2862 } else { 2863 // Remove the new phi from the graph and use the hit 2864 _igvn.remove_dead_node(phi); 2865 phi = hit; 2866 } 2867 _igvn.replace_input_of(use, idx, phi); 2868 } 2869 2870 #ifdef ASSERT 2871 //------------------------------ is_valid_loop_partition ------------------------------------- 2872 // Validate the loop partition sets: peel and not_peel 2873 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2874 VectorSet& not_peel ) { 2875 uint i; 2876 // Check that peel_list entries are in the peel set 2877 for (i = 0; i < peel_list.size(); i++) { 2878 if (!peel.test(peel_list.at(i)->_idx)) { 2879 return false; 2880 } 2881 } 2882 // Check at loop members are in one of peel set or not_peel set 2883 for (i = 0; i < loop->_body.size(); i++ ) { 2884 Node *def = loop->_body.at(i); 2885 uint di = def->_idx; 2886 // Check that peel set elements are in peel_list 2887 if (peel.test(di)) { 2888 if (not_peel.test(di)) { 2889 return false; 2890 } 2891 // Must be in peel_list also 2892 bool found = false; 2893 for (uint j = 0; j < peel_list.size(); j++) { 2894 if (peel_list.at(j)->_idx == di) { 2895 found = true; 2896 break; 2897 } 2898 } 2899 if (!found) { 2900 return false; 2901 } 2902 } else if (not_peel.test(di)) { 2903 if (peel.test(di)) { 2904 return false; 2905 } 2906 } else { 2907 return false; 2908 } 2909 } 2910 return true; 2911 } 2912 2913 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2914 // Ensure a use outside of loop is of the right form 2915 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2916 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2917 return (use->is_Phi() && 2918 use_c->is_Region() && use_c->req() == 3 && 2919 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2920 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2921 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2922 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2923 } 2924 2925 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2926 // Ensure that all uses outside of loop are of the right form 2927 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2928 uint orig_exit_idx, uint clone_exit_idx) { 2929 uint len = peel_list.size(); 2930 for (uint i = 0; i < len; i++) { 2931 Node *def = peel_list.at(i); 2932 2933 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2934 Node *use = def->fast_out(j); 2935 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2936 if (!loop->is_member(get_loop(use_c))) { 2937 // use is not in the loop, check for correct structure 2938 if (use->in(0) == def) { 2939 // Okay 2940 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2941 return false; 2942 } 2943 } 2944 } 2945 } 2946 return true; 2947 } 2948 #endif 2949 2950 //------------------------------ partial_peel ------------------------------------- 2951 // Partially peel (aka loop rotation) the top portion of a loop (called 2952 // the peel section below) by cloning it and placing one copy just before 2953 // the new loop head and the other copy at the bottom of the new loop. 2954 // 2955 // before after where it came from 2956 // 2957 // stmt1 stmt1 2958 // loop: stmt2 clone 2959 // stmt2 if condA goto exitA clone 2960 // if condA goto exitA new_loop: new 2961 // stmt3 stmt3 clone 2962 // if !condB goto loop if condB goto exitB clone 2963 // exitB: stmt2 orig 2964 // stmt4 if !condA goto new_loop orig 2965 // exitA: goto exitA 2966 // exitB: 2967 // stmt4 2968 // exitA: 2969 // 2970 // Step 1: find the cut point: an exit test on probable 2971 // induction variable. 2972 // Step 2: schedule (with cloning) operations in the peel 2973 // section that can be executed after the cut into 2974 // the section that is not peeled. This may need 2975 // to clone operations into exit blocks. For 2976 // instance, a reference to A[i] in the not-peel 2977 // section and a reference to B[i] in an exit block 2978 // may cause a left-shift of i by 2 to be placed 2979 // in the peel block. This step will clone the left 2980 // shift into the exit block and sink the left shift 2981 // from the peel to the not-peel section. 2982 // Step 3: clone the loop, retarget the control, and insert 2983 // phis for values that are live across the new loop 2984 // head. This is very dependent on the graph structure 2985 // from clone_loop. It creates region nodes for 2986 // exit control and associated phi nodes for values 2987 // flow out of the loop through that exit. The region 2988 // node is dominated by the clone's control projection. 2989 // So the clone's peel section is placed before the 2990 // new loop head, and the clone's not-peel section is 2991 // forms the top part of the new loop. The original 2992 // peel section forms the tail of the new loop. 2993 // Step 4: update the dominator tree and recompute the 2994 // dominator depth. 2995 // 2996 // orig 2997 // 2998 // stmt1 2999 // | 3000 // v 3001 // loop predicate 3002 // | 3003 // v 3004 // loop<----+ 3005 // | | 3006 // stmt2 | 3007 // | | 3008 // v | 3009 // ifA | 3010 // / | | 3011 // v v | 3012 // false true ^ <-- last_peel 3013 // / | | 3014 // / ===|==cut | 3015 // / stmt3 | <-- first_not_peel 3016 // / | | 3017 // | v | 3018 // v ifB | 3019 // exitA: / \ | 3020 // / \ | 3021 // v v | 3022 // false true | 3023 // / \ | 3024 // / ----+ 3025 // | 3026 // v 3027 // exitB: 3028 // stmt4 3029 // 3030 // 3031 // after clone loop 3032 // 3033 // stmt1 3034 // | 3035 // v 3036 // loop predicate 3037 // / \ 3038 // clone / \ orig 3039 // / \ 3040 // / \ 3041 // v v 3042 // +---->loop loop<----+ 3043 // | | | | 3044 // | stmt2 stmt2 | 3045 // | | | | 3046 // | v v | 3047 // | ifA ifA | 3048 // | | \ / | | 3049 // | v v v v | 3050 // ^ true false false true ^ <-- last_peel 3051 // | | ^ \ / | | 3052 // | cut==|== \ \ / ===|==cut | 3053 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3054 // | | dom | | | | 3055 // | v \ 1v v2 v | 3056 // | ifB regionA ifB | 3057 // | / \ | / \ | 3058 // | / \ v / \ | 3059 // | v v exitA: v v | 3060 // | true false false true | 3061 // | / ^ \ / \ | 3062 // +---- \ \ / ----+ 3063 // dom \ / 3064 // \ 1v v2 3065 // regionB 3066 // | 3067 // v 3068 // exitB: 3069 // stmt4 3070 // 3071 // 3072 // after partial peel 3073 // 3074 // stmt1 3075 // | 3076 // v 3077 // loop predicate 3078 // / 3079 // clone / orig 3080 // / TOP 3081 // / \ 3082 // v v 3083 // TOP->loop loop----+ 3084 // | | | 3085 // stmt2 stmt2 | 3086 // | | | 3087 // v v | 3088 // ifA ifA | 3089 // | \ / | | 3090 // v v v v | 3091 // true false false true | <-- last_peel 3092 // | ^ \ / +------|---+ 3093 // +->newloop \ \ / === ==cut | | 3094 // | stmt3 \ \ / TOP | | 3095 // | | dom | | stmt3 | | <-- first_not_peel 3096 // | v \ 1v v2 v | | 3097 // | ifB regionA ifB ^ v 3098 // | / \ | / \ | | 3099 // | / \ v / \ | | 3100 // | v v exitA: v v | | 3101 // | true false false true | | 3102 // | / ^ \ / \ | | 3103 // | | \ \ / v | | 3104 // | | dom \ / TOP | | 3105 // | | \ 1v v2 | | 3106 // ^ v regionB | | 3107 // | | | | | 3108 // | | v ^ v 3109 // | | exitB: | | 3110 // | | stmt4 | | 3111 // | +------------>-----------------+ | 3112 // | | 3113 // +-----------------<---------------------+ 3114 // 3115 // 3116 // final graph 3117 // 3118 // stmt1 3119 // | 3120 // v 3121 // loop predicate 3122 // | 3123 // v 3124 // stmt2 clone 3125 // | 3126 // v 3127 // ........> ifA clone 3128 // : / | 3129 // dom / | 3130 // : v v 3131 // : false true 3132 // : | | 3133 // : | v 3134 // : | newloop<-----+ 3135 // : | | | 3136 // : | stmt3 clone | 3137 // : | | | 3138 // : | v | 3139 // : | ifB | 3140 // : | / \ | 3141 // : | v v | 3142 // : | false true | 3143 // : | | | | 3144 // : | v stmt2 | 3145 // : | exitB: | | 3146 // : | stmt4 v | 3147 // : | ifA orig | 3148 // : | / \ | 3149 // : | / \ | 3150 // : | v v | 3151 // : | false true | 3152 // : | / \ | 3153 // : v v -----+ 3154 // RegionA 3155 // | 3156 // v 3157 // exitA 3158 // 3159 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3160 3161 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3162 if (!loop->_head->is_Loop()) { 3163 return false; 3164 } 3165 LoopNode *head = loop->_head->as_Loop(); 3166 3167 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3168 return false; 3169 } 3170 3171 // Check for complex exit control 3172 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3173 Node *n = loop->_body.at(ii); 3174 int opc = n->Opcode(); 3175 if (n->is_Call() || 3176 opc == Op_Catch || 3177 opc == Op_CatchProj || 3178 opc == Op_Jump || 3179 opc == Op_JumpProj) { 3180 #ifndef PRODUCT 3181 if (TracePartialPeeling) { 3182 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3183 } 3184 #endif 3185 return false; 3186 } 3187 } 3188 3189 int dd = dom_depth(head); 3190 3191 // Step 1: find cut point 3192 3193 // Walk up dominators to loop head looking for first loop exit 3194 // which is executed on every path thru loop. 3195 IfNode *peel_if = NULL; 3196 IfNode *peel_if_cmpu = NULL; 3197 3198 Node *iff = loop->tail(); 3199 while (iff != head) { 3200 if (iff->is_If()) { 3201 Node *ctrl = get_ctrl(iff->in(1)); 3202 if (ctrl->is_top()) return false; // Dead test on live IF. 3203 // If loop-varying exit-test, check for induction variable 3204 if (loop->is_member(get_loop(ctrl)) && 3205 loop->is_loop_exit(iff) && 3206 is_possible_iv_test(iff)) { 3207 Node* cmp = iff->in(1)->in(1); 3208 if (cmp->Opcode() == Op_CmpI) { 3209 peel_if = iff->as_If(); 3210 } else { 3211 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3212 peel_if_cmpu = iff->as_If(); 3213 } 3214 } 3215 } 3216 iff = idom(iff); 3217 } 3218 3219 // Prefer signed compare over unsigned compare. 3220 IfNode* new_peel_if = NULL; 3221 if (peel_if == NULL) { 3222 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3223 return false; // No peel point found 3224 } 3225 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3226 if (new_peel_if == NULL) { 3227 return false; // No peel point found 3228 } 3229 peel_if = new_peel_if; 3230 } 3231 Node* last_peel = stay_in_loop(peel_if, loop); 3232 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3233 if (first_not_peeled == NULL || first_not_peeled == head) { 3234 return false; 3235 } 3236 3237 #ifndef PRODUCT 3238 if (TraceLoopOpts) { 3239 tty->print("PartialPeel "); 3240 loop->dump_head(); 3241 } 3242 3243 if (TracePartialPeeling) { 3244 tty->print_cr("before partial peel one iteration"); 3245 Node_List wl; 3246 Node* t = head->in(2); 3247 while (true) { 3248 wl.push(t); 3249 if (t == head) break; 3250 t = idom(t); 3251 } 3252 while (wl.size() > 0) { 3253 Node* tt = wl.pop(); 3254 tt->dump(); 3255 if (tt == last_peel) tty->print_cr("-- cut --"); 3256 } 3257 } 3258 #endif 3259 ResourceArea *area = Thread::current()->resource_area(); 3260 VectorSet peel(area); 3261 VectorSet not_peel(area); 3262 Node_List peel_list(area); 3263 Node_List worklist(area); 3264 Node_List sink_list(area); 3265 3266 if (!may_require_nodes(loop->est_loop_clone_sz(2))) { 3267 return false; 3268 } 3269 3270 // Set of cfg nodes to peel are those that are executable from 3271 // the head through last_peel. 3272 assert(worklist.size() == 0, "should be empty"); 3273 worklist.push(head); 3274 peel.set(head->_idx); 3275 while (worklist.size() > 0) { 3276 Node *n = worklist.pop(); 3277 if (n != last_peel) { 3278 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3279 Node* use = n->fast_out(j); 3280 if (use->is_CFG() && 3281 loop->is_member(get_loop(use)) && 3282 !peel.test_set(use->_idx)) { 3283 worklist.push(use); 3284 } 3285 } 3286 } 3287 } 3288 3289 // Set of non-cfg nodes to peel are those that are control 3290 // dependent on the cfg nodes. 3291 uint i; 3292 for(i = 0; i < loop->_body.size(); i++ ) { 3293 Node *n = loop->_body.at(i); 3294 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3295 if (peel.test(n_c->_idx)) { 3296 peel.set(n->_idx); 3297 } else { 3298 not_peel.set(n->_idx); 3299 } 3300 } 3301 3302 // Step 2: move operations from the peeled section down into the 3303 // not-peeled section 3304 3305 // Get a post order schedule of nodes in the peel region 3306 // Result in right-most operand. 3307 scheduled_nodelist(loop, peel, peel_list ); 3308 3309 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3310 3311 // For future check for too many new phis 3312 uint old_phi_cnt = 0; 3313 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3314 Node* use = head->fast_out(j); 3315 if (use->is_Phi()) old_phi_cnt++; 3316 } 3317 3318 #ifndef PRODUCT 3319 if (TracePartialPeeling) { 3320 tty->print_cr("\npeeled list"); 3321 } 3322 #endif 3323 3324 // Evacuate nodes in peel region into the not_peeled region if possible 3325 uint new_phi_cnt = 0; 3326 uint cloned_for_outside_use = 0; 3327 for (i = 0; i < peel_list.size();) { 3328 Node* n = peel_list.at(i); 3329 #ifndef PRODUCT 3330 if (TracePartialPeeling) n->dump(); 3331 #endif 3332 bool incr = true; 3333 if ( !n->is_CFG() ) { 3334 3335 if ( has_use_in_set(n, not_peel) ) { 3336 3337 // If not used internal to the peeled region, 3338 // move "n" from peeled to not_peeled region. 3339 3340 if ( !has_use_internal_to_set(n, peel, loop) ) { 3341 3342 // if not pinned and not a load (which maybe anti-dependent on a store) 3343 // and not a CMove (Matcher expects only bool->cmove). 3344 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3345 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3346 sink_list.push(n); 3347 peel >>= n->_idx; // delete n from peel set. 3348 not_peel <<= n->_idx; // add n to not_peel set. 3349 peel_list.remove(i); 3350 incr = false; 3351 #ifndef PRODUCT 3352 if (TracePartialPeeling) { 3353 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3354 n->_idx, get_ctrl(n)->_idx); 3355 } 3356 #endif 3357 } 3358 } else { 3359 // Otherwise check for special def-use cases that span 3360 // the peel/not_peel boundary such as bool->if 3361 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3362 new_phi_cnt++; 3363 } 3364 } 3365 } 3366 if (incr) i++; 3367 } 3368 3369 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3370 #ifndef PRODUCT 3371 if (TracePartialPeeling) { 3372 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3373 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3374 } 3375 #endif 3376 if (new_peel_if != NULL) { 3377 remove_cmpi_loop_exit(new_peel_if, loop); 3378 } 3379 // Inhibit more partial peeling on this loop 3380 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3381 head->mark_partial_peel_failed(); 3382 if (cloned_for_outside_use > 0) { 3383 // Terminate this round of loop opts because 3384 // the graph outside this loop was changed. 3385 C->set_major_progress(); 3386 return true; 3387 } 3388 return false; 3389 } 3390 3391 // Step 3: clone loop, retarget control, and insert new phis 3392 3393 // Create new loop head for new phis and to hang 3394 // the nodes being moved (sinked) from the peel region. 3395 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3396 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3397 _igvn.register_new_node_with_optimizer(new_head); 3398 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3399 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3400 set_loop(new_head, loop); 3401 loop->_body.push(new_head); 3402 not_peel.set(new_head->_idx); 3403 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3404 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3405 3406 while (sink_list.size() > 0) { 3407 Node* n = sink_list.pop(); 3408 set_ctrl(n, new_head); 3409 } 3410 3411 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3412 3413 clone_loop(loop, old_new, dd, IgnoreStripMined); 3414 3415 const uint clone_exit_idx = 1; 3416 const uint orig_exit_idx = 2; 3417 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3418 3419 Node* head_clone = old_new[head->_idx]; 3420 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3421 Node* orig_tail_clone = head_clone->in(2); 3422 3423 // Add phi if "def" node is in peel set and "use" is not 3424 3425 for(i = 0; i < peel_list.size(); i++ ) { 3426 Node *def = peel_list.at(i); 3427 if (!def->is_CFG()) { 3428 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3429 Node *use = def->fast_out(j); 3430 if (has_node(use) && use->in(0) != C->top() && 3431 (!peel.test(use->_idx) || 3432 (use->is_Phi() && use->in(0) == head)) ) { 3433 worklist.push(use); 3434 } 3435 } 3436 while( worklist.size() ) { 3437 Node *use = worklist.pop(); 3438 for (uint j = 1; j < use->req(); j++) { 3439 Node* n = use->in(j); 3440 if (n == def) { 3441 3442 // "def" is in peel set, "use" is not in peel set 3443 // or "use" is in the entry boundary (a phi) of the peel set 3444 3445 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3446 3447 if ( loop->is_member(get_loop( use_c )) ) { 3448 // use is in loop 3449 if (old_new[use->_idx] != NULL) { // null for dead code 3450 Node* use_clone = old_new[use->_idx]; 3451 _igvn.replace_input_of(use, j, C->top()); 3452 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3453 } 3454 } else { 3455 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3456 // use is not in the loop, check if the live range includes the cut 3457 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3458 if (not_peel.test(lp_if->_idx)) { 3459 assert(j == orig_exit_idx, "use from original loop"); 3460 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3461 } 3462 } 3463 } 3464 } 3465 } 3466 } 3467 } 3468 3469 // Step 3b: retarget control 3470 3471 // Redirect control to the new loop head if a cloned node in 3472 // the not_peeled region has control that points into the peeled region. 3473 // This necessary because the cloned peeled region will be outside 3474 // the loop. 3475 // from to 3476 // cloned-peeled <---+ 3477 // new_head_clone: | <--+ 3478 // cloned-not_peeled in(0) in(0) 3479 // orig-peeled 3480 3481 for(i = 0; i < loop->_body.size(); i++ ) { 3482 Node *n = loop->_body.at(i); 3483 if (!n->is_CFG() && n->in(0) != NULL && 3484 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3485 Node* n_clone = old_new[n->_idx]; 3486 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3487 } 3488 } 3489 3490 // Backedge of the surviving new_head (the clone) is original last_peel 3491 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3492 3493 // Cut first node in original not_peel set 3494 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3495 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3496 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3497 3498 // Copy head_clone back-branch info to original head 3499 // and remove original head's loop entry and 3500 // clone head's back-branch 3501 _igvn.rehash_node_delayed(head); // Multiple edge updates 3502 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3503 head->set_req(LoopNode::LoopBackControl, C->top()); 3504 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3505 3506 // Similarly modify the phis 3507 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3508 Node* use = head->fast_out(k); 3509 if (use->is_Phi() && use->outcnt() > 0) { 3510 Node* use_clone = old_new[use->_idx]; 3511 _igvn.rehash_node_delayed(use); // Multiple edge updates 3512 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3513 use->set_req(LoopNode::LoopBackControl, C->top()); 3514 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3515 } 3516 } 3517 3518 // Step 4: update dominator tree and dominator depth 3519 3520 set_idom(head, orig_tail_clone, dd); 3521 recompute_dom_depth(); 3522 3523 // Inhibit more partial peeling on this loop 3524 new_head_clone->set_partial_peel_loop(); 3525 C->set_major_progress(); 3526 loop->record_for_igvn(); 3527 3528 #ifndef PRODUCT 3529 if (TracePartialPeeling) { 3530 tty->print_cr("\nafter partial peel one iteration"); 3531 Node_List wl(area); 3532 Node* t = last_peel; 3533 while (true) { 3534 wl.push(t); 3535 if (t == head_clone) break; 3536 t = idom(t); 3537 } 3538 while (wl.size() > 0) { 3539 Node* tt = wl.pop(); 3540 if (tt == head) tty->print_cr("orig head"); 3541 else if (tt == new_head_clone) tty->print_cr("new head"); 3542 else if (tt == head_clone) tty->print_cr("clone head"); 3543 tt->dump(); 3544 } 3545 } 3546 #endif 3547 return true; 3548 } 3549 3550 //------------------------------reorg_offsets---------------------------------- 3551 // Reorganize offset computations to lower register pressure. Mostly 3552 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3553 // then alive with the post-incremented trip counter forcing an extra 3554 // register move) 3555 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3556 // Perform it only for canonical counted loops. 3557 // Loop's shape could be messed up by iteration_split_impl. 3558 if (!loop->_head->is_CountedLoop()) 3559 return; 3560 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3561 return; 3562 3563 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3564 CountedLoopEndNode *cle = cl->loopexit(); 3565 Node *exit = cle->proj_out(false); 3566 Node *phi = cl->phi(); 3567 3568 // Check for the special case when using the pre-incremented trip-counter on 3569 // the fall-out path (forces the pre-incremented and post-incremented trip 3570 // counter to be live at the same time). Fix this by adjusting to use the 3571 // post-increment trip counter. 3572 3573 bool progress = true; 3574 while (progress) { 3575 progress = false; 3576 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3577 Node* use = phi->fast_out(i); // User of trip-counter 3578 if (!has_ctrl(use)) continue; 3579 Node *u_ctrl = get_ctrl(use); 3580 if (use->is_Phi()) { 3581 u_ctrl = NULL; 3582 for (uint j = 1; j < use->req(); j++) 3583 if (use->in(j) == phi) 3584 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3585 } 3586 IdealLoopTree *u_loop = get_loop(u_ctrl); 3587 // Look for loop-invariant use 3588 if (u_loop == loop) continue; 3589 if (loop->is_member(u_loop)) continue; 3590 // Check that use is live out the bottom. Assuming the trip-counter 3591 // update is right at the bottom, uses of of the loop middle are ok. 3592 if (dom_lca(exit, u_ctrl) != exit) continue; 3593 // Hit! Refactor use to use the post-incremented tripcounter. 3594 // Compute a post-increment tripcounter. 3595 Node* c = exit; 3596 if (cl->is_strip_mined()) { 3597 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3598 if (!outer_loop->is_member(u_loop)) { 3599 c = cl->outer_loop_exit(); 3600 } 3601 } 3602 Node *opaq = new Opaque2Node(C, cle->incr()); 3603 register_new_node(opaq, c); 3604 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3605 set_ctrl(neg_stride, C->root()); 3606 Node *post = new AddINode(opaq, neg_stride); 3607 register_new_node(post, c); 3608 _igvn.rehash_node_delayed(use); 3609 for (uint j = 1; j < use->req(); j++) { 3610 if (use->in(j) == phi) 3611 use->set_req(j, post); 3612 } 3613 // Since DU info changed, rerun loop 3614 progress = true; 3615 break; 3616 } 3617 } 3618 3619 }